{ "site": "https://thecontinuitylayer.com", "publisher": "The Continuity Layer", "file": "https://thecontinuitylayer.com/open-data/tools.json", "about": "https://thecontinuitylayer.com/open-data", "license": "CC BY 4.0", "licenseUrl": "https://creativecommons.org/licenses/by/4.0/", "attribution": "The Continuity Layer, https://thecontinuitylayer.com", "generated": "2026-11-09T15:37:50.164Z", "method": "Prices and fees are the ones each vendor published on the day in the table; check the vendor before buying.", "categories": [ { "slug": "local-ai-apps", "name": "Local AI apps", "page": "https://thecontinuitylayer.com/tools/local-ai-apps", "description": "The apps that run open AI models on your own computer, compared in one table: Ollama, LM Studio, Jan, Open WebUI, AnythingLLM, llama.cpp, KoboldCpp, LocalAI and vLLM, with difficulty, license and our verdict.", "lastReviewed": "2026-11-06", "capabilityLabel": "Runs", "priceLabel": "Price", "entries": [ { "slug": "lm-studio", "name": "LM Studio", "maker": "LM Studio", "kind": "Desktop app", "page": "https://thecontinuitylayer.com/tools/local-ai-apps/lm-studio", "url": "https://lmstudio.ai/", "source": "https://lmstudio.ai/pricing", "price": "Free to run local models; paid plans for its agent and hosted models", "price_band": "Free; Bionic+ $20/month, Pro $100/month", "price_as_of": "2026-11-06", "capability": "GGUF models through llama.cpp, and MLX models on Apple Silicon", "forWho": "Beginners who want buttons, not a terminal", "best_for": "Easiest first install", "verdict": "The easiest start: a polished app that finds, downloads and runs models, with a local server when you want one. The app itself is not open source.", "description": [ "LM Studio is a desktop app for finding, downloading and chatting with open models on your own computer. It runs models through llama.cpp, and through Apple's MLX on Apple Silicon Macs, so the same app works on a Mac or a PC.", "It is the gentlest way in: search for a model, click download, and chat. When you outgrow the chat window, it can serve the loaded model to other apps on your machine, and it has a command-line tool called lms. Its free plan runs local models; paid plans add LM Studio's own agent, Bionic, and hosted models. Hosted models run on LM Studio's servers, not yours." ], "specs": [ { "label": "Platforms", "value": "Apple Silicon Macs (macOS 14+), Windows x64 and ARM, Linux x64 and ARM64" }, { "label": "Memory", "value": "16GB recommended; 8GB Macs work with small models" }, { "label": "Difficulty", "value": "Easiest: no terminal" }, { "label": "Engines", "value": "llama.cpp and MLX" }, { "label": "Open source", "value": "No (the app's code is not published)" }, { "label": "Price", "value": "Free for local models; Bionic+ $20/month, Pro $100/month" } ], "goodFor": [ "A first local model in minutes", "Mac users who want MLX speed", "Browsing and comparing models" ], "notFor": [ "People who want only open-source software", "Running a shared server for a household or team" ], "sources": [ "https://lmstudio.ai/pricing", "https://lmstudio.ai/docs/app/system-requirements" ] }, { "slug": "ollama", "name": "Ollama", "maker": "Ollama", "kind": "Runner and local server", "page": "https://thecontinuitylayer.com/tools/local-ai-apps/ollama", "url": "https://ollama.com/", "source": "https://github.com/ollama/ollama", "price": "Free and open source (MIT); optional paid cloud models", "price_band": "Free (MIT license)", "price_as_of": "2026-11-06", "capability": "Its own library of hundreds of open models, one command each", "forWho": "Anyone who will build on a local model: apps, agents, chat front-ends", "best_for": "The standard base to build on", "verdict": "The default engine most other local apps plug into: one command to run a model, and an OpenAI-compatible API on your machine.", "description": [ "Ollama runs open models on your computer with one command, such as ollama run gemma4. It keeps a library of ready-made builds, so you don't hunt for files, and it runs a local server that other programs talk to.", "That server speaks a subset of the OpenAI API at localhost:11434, which is why so many chat apps, coding tools and agents support Ollama out of the box. It also offers cloud models with names ending in :cloud; those run on Ollama's servers and need an account, so skip them if everything must stay on your machine." ], "specs": [ { "label": "Platforms", "value": "macOS, Windows, Linux, Docker" }, { "label": "Difficulty", "value": "Easy: a few terminal commands" }, { "label": "API", "value": "Own REST API plus an OpenAI-compatible subset at localhost:11434/v1" }, { "label": "License", "value": "MIT" }, { "label": "Cloud", "value": "Optional :cloud models run on Ollama's servers" } ], "goodFor": [ "Connecting chat apps, coding tools and agents to a local model", "Switching models quickly", "Scripting and automation" ], "notFor": [ "People who never want to open a terminal (use LM Studio or Jan)", "Serving many users at once (use vLLM)" ], "sources": [ "https://github.com/ollama/ollama", "https://docs.ollama.com/api/openai-compatibility", "https://ollama.com/library" ] }, { "slug": "jan", "name": "Jan", "maker": "Jan (janhq)", "kind": "Desktop app", "page": "https://thecontinuitylayer.com/tools/local-ai-apps/jan", "url": "https://jan.ai/", "source": "https://github.com/janhq/jan", "price": "Free and open source (Apache 2.0)", "price_band": "Free (Apache 2.0)", "price_as_of": "2026-11-06", "capability": "Open models from Hugging Face through bundled llama.cpp; optional hosted models", "forWho": "People who want an open-source ChatGPT-style app that works offline", "best_for": "Open-source desktop chat", "verdict": "The open-source answer to LM Studio: a clean chat app that runs models offline, with a local API and MCP tools.", "description": [ "Jan is an open-source desktop app that works as a local alternative to ChatGPT. It bundles llama.cpp, downloads open models such as Gemma, Qwen and gpt-oss, and can run entirely offline.", "It can also connect to hosted models from OpenAI, Anthropic, Mistral and others, so one window covers both. It serves an OpenAI-compatible API at localhost:1337 and supports MCP for tool use." ], "specs": [ { "label": "Platforms", "value": "Windows 10+, macOS 13.6+, Linux" }, { "label": "Memory", "value": "Maker's guide: 8GB for 3B models, 16GB for 7B, 32GB for 13B" }, { "label": "Difficulty", "value": "Easy: no terminal" }, { "label": "API", "value": "OpenAI-compatible at localhost:1337" }, { "label": "License", "value": "Apache 2.0" } ], "goodFor": [ "Offline chat with open-source software only", "Mixing local and hosted models in one app", "MCP tools" ], "notFor": [ "Households or teams sharing one server (use Open WebUI)" ], "sources": [ "https://github.com/janhq/jan" ] }, { "slug": "open-webui", "name": "Open WebUI", "maker": "Open WebUI", "kind": "Self-hosted chat interface", "page": "https://thecontinuitylayer.com/tools/local-ai-apps/open-webui", "url": "https://openwebui.com/", "source": "https://github.com/open-webui/open-webui", "price": "Free to self-host (Open WebUI License)", "price_band": "Free to self-host", "price_as_of": "2026-11-06", "capability": "Any Ollama or OpenAI-compatible backend, local or hosted", "forWho": "Households, teams and tinkerers who want their own ChatGPT in a browser", "best_for": "Your own ChatGPT for the house", "verdict": "The most complete front end: accounts, documents, memory and web search on top of Ollama or any OpenAI-style API.", "description": [ "Open WebUI is a self-hosted web interface that looks and works like ChatGPT, on top of Ollama or any OpenAI-compatible server. You install it once and everyone on your network can use it from a browser.", "It adds what a bare model lacks: chat with your documents (retrieval across a choice of vector databases), a memory that carries facts across conversations, web search through many providers, and user accounts with roles. It can run fully offline. Its license, the Open WebUI License, requires keeping the Open WebUI branding." ], "specs": [ { "label": "Install", "value": "pip install open-webui (Python 3.12 recommended), Docker, Kubernetes" }, { "label": "Difficulty", "value": "Medium: Docker or Python" }, { "label": "Backends", "value": "Ollama and any OpenAI-compatible API" }, { "label": "Features", "value": "Document chat, memory, web search, accounts and roles" }, { "label": "License", "value": "Open WebUI License (keep the branding)" } ], "goodFor": [ "A shared AI for a family or small team", "Chatting with your own documents", "One window over local and hosted models" ], "notFor": [ "Someone who wants a one-click desktop app (use LM Studio or Jan)", "Rebranding it as your own product" ], "sources": [ "https://github.com/open-webui/open-webui" ] }, { "slug": "anythingllm", "name": "AnythingLLM", "maker": "Mintplex Labs", "kind": "Desktop app and server", "page": "https://thecontinuitylayer.com/tools/local-ai-apps/anythingllm", "url": "https://anythingllm.com/", "source": "https://github.com/Mintplex-Labs/anything-llm", "price": "Free and open source (MIT); a hosted instance is sold", "price_band": "Free (MIT); hosted price not published", "price_as_of": "2026-11-06", "capability": "Built-in local engine, or Ollama, LM Studio, LocalAI and hosted APIs", "forWho": "People who mainly want to chat with their documents and run simple agents", "best_for": "Chat with your documents", "verdict": "The quickest route to a private assistant over your own files, with agents built in; multi-user needs the Docker version.", "description": [ "AnythingLLM is an all-in-one app for a private ChatGPT-style assistant built around your documents. Drop in PDFs, Word files and text, and it answers with citations to the source.", "It has a built-in local engine and also works with Ollama, LM Studio and LocalAI, or hosted providers. Agents can browse the web and run scheduled tasks, and it supports MCP. The desktop app is single-user; accounts, permissions and an embeddable chat widget are in the Docker version." ], "specs": [ { "label": "Platforms", "value": "Desktop for Mac, Windows, Linux; Docker; cloud templates" }, { "label": "Difficulty", "value": "Easy (desktop) to medium (Docker)" }, { "label": "Documents", "value": "PDF, DOCX, TXT and more, with citations" }, { "label": "Agents", "value": "Built-in agents, no-code builder, scheduled tasks, MCP" }, { "label": "License", "value": "MIT" } ], "goodFor": [ "A private assistant over a folder of files", "Simple agents without code" ], "notFor": [ "People who only want plain chat (LM Studio or Jan are simpler)" ], "sources": [ "https://github.com/Mintplex-Labs/anything-llm" ] }, { "slug": "llama-cpp", "name": "llama.cpp", "maker": "ggml-org", "kind": "Inference engine", "page": "https://thecontinuitylayer.com/tools/local-ai-apps/llama-cpp", "url": "https://github.com/ggml-org/llama.cpp", "source": "https://github.com/ggml-org/llama.cpp", "price": "Free and open source (MIT)", "price_band": "Free (MIT license)", "price_as_of": "2026-11-06", "capability": "GGUF models on almost any hardware: CPU, NVIDIA, Apple, AMD, Intel, Vulkan", "forWho": "Tinkerers who want the most control and the newest features first", "best_for": "Maximum control, any hardware", "verdict": "The engine under much of local AI. More setup than Ollama, but every setting is yours, and new models often land here first.", "description": [ "llama.cpp is the C/C++ engine that made running models on ordinary computers practical. It runs GGUF model files on processors and graphics cards from every major maker, with no other dependencies.", "LM Studio, Jan, KoboldCpp and others build on it. Used directly, it gives you a server with an OpenAI-compatible API and a built-in web chat, and full control over every setting. The trade-off is more to learn: you pick the files and the flags yourself." ], "specs": [ { "label": "Hardware", "value": "CPU (x86, ARM), CUDA, Metal, Vulkan, SYCL, HIP and more" }, { "label": "Difficulty", "value": "Advanced: terminal and settings" }, { "label": "Format", "value": "GGUF" }, { "label": "API", "value": "OpenAI-compatible server with a web UI" }, { "label": "License", "value": "MIT" } ], "goodFor": [ "Squeezing the most out of unusual or older hardware", "Trying new models the day they land" ], "notFor": [ "Beginners (start with LM Studio or Ollama)" ], "sources": [ "https://github.com/ggml-org/llama.cpp" ] }, { "slug": "koboldcpp", "name": "KoboldCpp", "maker": "LostRuins", "kind": "Single-file runner", "page": "https://thecontinuitylayer.com/tools/local-ai-apps/koboldcpp", "url": "https://github.com/LostRuins/koboldcpp", "source": "https://github.com/LostRuins/koboldcpp", "price": "Free and open source (AGPL-3.0)", "price_band": "Free (AGPL-3.0)", "price_as_of": "2026-11-06", "capability": "GGUF text models plus image generation, built on llama.cpp", "forWho": "Fiction writers and role-players", "best_for": "Stories and role-play", "verdict": "One file, no install, with writing tools (memory, world info, character cards) the general apps don't have.", "description": [ "KoboldCpp is a single executable that runs GGUF models with a built-in web interface, KoboldAI Lite. There is nothing to install: download one file and run it.", "It is built for long-form writing and role-play, with memory, world info, author's notes and support for imported character cards. It can also generate images with Stable Diffusion and Flux models, and it serves an OpenAI-compatible API alongside its own." ], "specs": [ { "label": "Platforms", "value": "Windows, Linux, Apple Silicon Mac; more from source" }, { "label": "Difficulty", "value": "Easy to start, many settings" }, { "label": "API", "value": "OpenAI-compatible at /v1, KoboldAI API, others" }, { "label": "License", "value": "AGPL-3.0" } ], "goodFor": [ "Fiction and role-play", "A portable setup on a USB stick" ], "notFor": [ "Document chat or business use (use AnythingLLM or Open WebUI)" ], "sources": [ "https://github.com/LostRuins/koboldcpp" ] }, { "slug": "localai", "name": "LocalAI", "maker": "LocalAI", "kind": "Self-hosted API server", "page": "https://thecontinuitylayer.com/tools/local-ai-apps/localai", "url": "https://localai.io/", "source": "https://github.com/mudler/LocalAI", "price": "Free and open source (MIT)", "price_band": "Free (MIT license)", "price_as_of": "2026-11-06", "capability": "Text, images, voice, embeddings and video behind OpenAI-style APIs", "forWho": "Developers replacing paid AI APIs with their own server", "best_for": "Drop-in replacement for AI APIs", "verdict": "One server that answers like OpenAI's, Anthropic's and ElevenLabs' APIs, for apps you can't rewrite; runs without a GPU.", "description": [ "LocalAI is a self-hosted server that copies the shape of popular AI APIs, including OpenAI's, Anthropic's and ElevenLabs', so existing apps can point at it instead of a paid service.", "It covers more than chat: image generation, speech to text, text to speech, embeddings and more. It runs on a processor alone or with NVIDIA, AMD, Intel or Apple graphics, from Docker or a single binary." ], "specs": [ { "label": "Install", "value": "Docker, macOS app, binary, Kubernetes" }, { "label": "Difficulty", "value": "Medium: Docker and config" }, { "label": "Hardware", "value": "CPU-only, or CUDA, ROCm, oneAPI, Metal, Vulkan" }, { "label": "License", "value": "MIT" } ], "goodFor": [ "Swapping a paid API for a local one in an existing app", "Voice and images as well as text" ], "notFor": [ "Someone who just wants to chat (use LM Studio, Jan or Ollama)" ], "sources": [ "https://github.com/mudler/LocalAI" ] }, { "slug": "vllm", "name": "vLLM", "maker": "vLLM project", "kind": "Serving engine", "page": "https://thecontinuitylayer.com/tools/local-ai-apps/vllm", "url": "https://docs.vllm.ai/", "source": "https://github.com/vllm-project/vllm", "price": "Free and open source (Apache 2.0)", "price_band": "Free (Apache 2.0)", "price_as_of": "2026-11-06", "capability": "Most open models on NVIDIA, AMD and Intel GPUs, plus TPUs and others by plugin", "forWho": "Teams serving a model to many users at once", "best_for": "Serving many users", "verdict": "The serving engine for real traffic on rented or owned GPUs; overkill for one person on a laptop.", "description": [ "vLLM is an engine for serving language models fast, to many people at once. It is what you run on a rented GPU server when a model has to answer a team, an app or a business, not one person.", "It offers an OpenAI-compatible server, plus Anthropic Messages API and gRPC support, and runs on NVIDIA, AMD and Intel graphics cards, with plugins for Google TPUs, Apple Silicon and others. It is installed with pip and configured from the command line." ], "specs": [ { "label": "Hardware", "value": "NVIDIA, AMD, Intel GPUs; TPU, Apple Silicon and others by plugin" }, { "label": "Difficulty", "value": "Advanced: Linux, Python, GPU drivers" }, { "label": "API", "value": "OpenAI-compatible, Anthropic Messages, gRPC" }, { "label": "License", "value": "Apache 2.0" } ], "goodFor": [ "Many users on one model", "Self-hosting on a rented GPU" ], "notFor": [ "Personal use on a laptop (use Ollama or LM Studio)" ], "sources": [ "https://github.com/vllm-project/vllm" ] } ] }, { "slug": "local-hardware", "name": "Local hardware", "page": "https://thecontinuitylayer.com/tools/local-hardware", "description": "Graphics cards, Macs and 128GB mini PCs for running open AI models at home, compared in one table: price band, what size of model each runs, and our one-line verdict.", "lastReviewed": "2026-10-31", "capabilityLabel": "Runs", "priceLabel": "Price", "entries": [ { "slug": "nvidia-geforce-rtx-5060-ti-16gb", "name": "GeForce RTX 5060 Ti 16GB", "maker": "NVIDIA", "kind": "GPU", "page": "https://thecontinuitylayer.com/tools/local-hardware/nvidia-geforce-rtx-5060-ti-16gb", "url": "https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5060-family/", "source": "https://www.nvidia.com/en-us/geforce/news/rtx-5060-desktop-family-laptop-5060-coming-soon/", "price": "$429 (NVIDIA starting price; retail prices vary)", "price_band": "$429 (16GB, NVIDIA starting price)", "price_as_of": "2026-11-02", "capability": "Up to ~14B models at 4-bit; ~24B fits with short context", "forWho": "People adding a first local-AI card to an existing desktop on a budget", "best_for": "First local-AI card on a budget", "verdict": "The cheapest new 16GB card: fine for 8–14B models, slow and tight above that.", "description": [ "The RTX 5060 Ti 16GB is NVIDIA's cheapest current card with 16GB of video memory. It drops into an ordinary desktop PC and works with every common local-AI app, from Ollama to LM Studio.", "16GB is enough for 8–14B models with room for a decent context window, and small 20–24B models at tight settings. It is noticeably slower than 24GB-and-up cards on the same model. Retail prices vary by board maker and store, and many listings sit above NVIDIA's starting price." ], "specs": [ { "label": "Memory", "value": "16GB GDDR7" }, { "label": "Interface", "value": "PCIe 5.0" }, { "label": "Software", "value": "CUDA; supported by all major local-AI apps" } ], "goodFor": [ "8–14B chat and coding models", "Image generation at modest sizes", "Learning local AI cheaply" ], "notFor": [ "30B-plus models", "Long-context agent work" ], "sources": [ "https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5060-family/", "https://www.nvidia.com/en-us/geforce/news/rtx-5060-desktop-family-laptop-5060-coming-soon/" ] }, { "slug": "nvidia-geforce-rtx-5090", "name": "GeForce RTX 5090", "maker": "NVIDIA", "kind": "GPU", "page": "https://thecontinuitylayer.com/tools/local-hardware/nvidia-geforce-rtx-5090", "url": "https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/", "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/", "price": "$1,999 (NVIDIA starting price, Founders Edition; partner cards cost more)", "price_band": "", "price_as_of": "", "capability": "Up to ~32B models at 8-bit, ~70B only with heavy quantization or CPU offload", "forWho": "Enthusiasts who want the fastest single consumer card and can absorb the price", "best_for": "Fastest single consumer card", "verdict": "The fastest card you can put in a desktop, at a price and power draw most people cannot justify.", "description": [ "The RTX 5090 is NVIDIA's top consumer graphics card, with 32GB of very fast memory. It is the fastest single card most people can buy for running open models at home.", "32GB fits 27–32B models comfortably with long context, and generation speed is the best of any card here. It draws a lot of power and needs a large case and power supply. Partner cards are listed well above the Founders Edition price, and retail prices vary." ], "specs": [ { "label": "Memory", "value": "32GB GDDR7" }, { "label": "Memory bandwidth", "value": "1,792 GB/s" }, { "label": "Board power", "value": "575W" }, { "label": "Interface", "value": "PCIe 5.0 x16" }, { "label": "Software", "value": "CUDA; best-supported platform" } ], "goodFor": [ "Fast 27–32B models with long context", "Mixing AI work with gaming", "Image and video generation" ], "notFor": [ "Buyers watching value", "70B-plus models at good quality", "Small or quiet builds" ], "sources": [ "https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/", "https://marketplace.nvidia.com/en-us/consumer/graphics-cards/geforce-rtx-5090-founders-edition/", "https://www.nvidia.com/en-us/geforce/graphics-cards/compare/" ] }, { "slug": "nvidia-geforce-rtx-3090-used", "name": "GeForce RTX 3090 (used)", "maker": "NVIDIA", "kind": "GPU (used)", "page": "https://thecontinuitylayer.com/tools/local-hardware/nvidia-geforce-rtx-3090-used", "url": "https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/", "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/", "price": "varies (used only)", "price_band": "", "price_as_of": "", "capability": "Up to ~32B models at 4-bit; 70B too slow on one card", "forWho": "Tinkerers who want 24GB of memory for the least money and are fine buying used", "best_for": "Cheapest 24GB with CUDA", "verdict": "Still the best value route to 24GB if you are happy buying used and checking the card.", "description": [ "The RTX 3090 is a 2020 card that is no longer made, but its 24GB of memory keeps it popular for local AI. It is bought second-hand, mostly on eBay.", "24GB runs 27–35B models at 4-bit at usable speed, though it is slower than an RTX 4090. Used prices vary widely, so check recent sold listings, test the card, and expect high power draw and heat." ], "specs": [ { "label": "Memory", "value": "24GB GDDR6X" }, { "label": "Memory bandwidth", "value": "936 GB/s" }, { "label": "Board power", "value": "350W" }, { "label": "Interface", "value": "PCIe 4.0 x16" }, { "label": "Condition", "value": "Used only; no maker warranty" } ], "goodFor": [ "Cheapest route to 24GB with CUDA", "Two-card 48GB builds", "27–32B models" ], "notFor": [ "Buyers who want a warranty", "Small or quiet PCs" ], "sources": [ "https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/", "https://www.nvidia.com/content/PDF/nvidia-ampere-ga-102-gpu-architecture-whitepaper-v2.1.pdf" ] }, { "slug": "nvidia-geforce-rtx-4090-used", "name": "GeForce RTX 4090 (used)", "maker": "NVIDIA", "kind": "GPU (used)", "page": "https://thecontinuitylayer.com/tools/local-hardware/nvidia-geforce-rtx-4090-used", "url": "https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/", "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/", "price": "varies (mostly used)", "price_band": "", "price_as_of": "", "capability": "Up to ~32B models at 4–8-bit; 70B too slow on one card", "forWho": "People who want fast 24GB performance without paying RTX 5090 prices", "best_for": "Fast 24GB without 5090 prices", "verdict": "Faster than a 3090 on the same models, but it fits no bigger ones, so pay the difference only for speed.", "description": [ "The RTX 4090 was NVIDIA's 2022 flagship and is now mainly sold used. It has 24GB of memory and is much faster than an RTX 3090 on the same models.", "It handles 27–32B models quickly and is a strong card for image generation. Used prices vary widely, which can narrow its value lead over a 3090. It is large and power-hungry; check the power connector on any used card." ], "specs": [ { "label": "Memory", "value": "24GB GDDR6X" }, { "label": "Memory bandwidth", "value": "1,008 GB/s" }, { "label": "Board power", "value": "450W" }, { "label": "Interface", "value": "PCIe 4.0" }, { "label": "Condition", "value": "Mostly used; warranty varies" } ], "goodFor": [ "Fast 24GB inference", "Image generation", "Mixed gaming and AI use" ], "notFor": [ "Value-first buyers (a 3090 fits the same models)", "Models above ~35B" ], "sources": [ "https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/", "https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvidia-ada-gpu-architecture.pdf" ] }, { "slug": "amd-radeon-ai-pro-r9700", "name": "AMD Radeon AI PRO R9700", "maker": "AMD", "kind": "GPU", "page": "https://thecontinuitylayer.com/tools/local-hardware/amd-radeon-ai-pro-r9700", "url": "https://www.amd.com/en/products/graphics/workstations/radeon-ai-pro/ai-9000-series/amd-radeon-ai-pro-r9700.html", "source": "https://www.amd.com/en/products/graphics/workstations/radeon-ai-pro/ai-9000-series/amd-radeon-ai-pro-r9700.html", "price": "$1,299 (AMD MSRP)", "price_band": "", "price_as_of": "", "capability": "Up to ~32B models at 4–8-bit; 70B only with heavy quantization", "forWho": "Linux-comfortable builders who want 32GB on one new card for less than an RTX 5090", "best_for": "32GB on one new card, Linux", "verdict": "The most memory on a single new card under the RTX 5090, if you can live without CUDA.", "description": [ "The Radeon AI PRO R9700 is AMD's workstation card with 32GB of memory. AMD lists an MSRP of $1,299.", "32GB gives the same model headroom as an RTX 5090 for less money, though it is slower. It runs llama.cpp, LM Studio and Ollama through ROCm or Vulkan, but some tools still assume NVIDIA. AMD also sells a passively cooled R9700S version for multi-card workstations." ], "specs": [ { "label": "Memory", "value": "32GB GDDR6" }, { "label": "Memory bandwidth", "value": "640 GB/s" }, { "label": "Board power", "value": "300W" }, { "label": "Software", "value": "ROCm, Vulkan" } ], "goodFor": [ "32GB on a single new card", "Multi-card workstations", "27–32B models with long context" ], "notFor": [ "People who need every CUDA-only tool", "Speed-first users" ], "sources": [ "https://www.amd.com/en/products/graphics/workstations/radeon-ai-pro/ai-9000-series/amd-radeon-ai-pro-r9700.html" ] }, { "slug": "intel-arc-pro-b60", "name": "Intel Arc Pro B60", "maker": "Intel", "kind": "GPU", "page": "https://thecontinuitylayer.com/tools/local-hardware/intel-arc-pro-b60", "url": "https://www.intel.com/content/www/us/en/products/sku/243916/intel-arc-pro-b60-graphics/specifications.html", "source": "https://www.intel.com/content/www/us/en/products/sku/243916/intel-arc-pro-b60-graphics/specifications.html", "price": "not published (sold through board makers; prices vary)", "price_band": "", "price_as_of": "", "capability": "Up to ~30B models at 4-bit; best at 14B", "forWho": "Budget builders who want 24GB new with a warranty and can handle setup", "best_for": "Cheap new 24GB, more setup", "verdict": "A low-cost 24GB card with a warranty, for people willing to do their own software setup.", "description": [ "The Arc Pro B60 is Intel's 24GB workstation card, sold through board makers such as ASRock and Sparkle. Intel does not publish a list price.", "Memory bandwidth is much lower than NVIDIA's 24GB cards, so it is slower on the same model. Software support is narrower: it works best with Intel's vLLM stack and Vulkan builds of llama.cpp. Intel also sells a 32GB Arc Pro B70." ], "specs": [ { "label": "Memory", "value": "24GB GDDR6" }, { "label": "Memory bandwidth", "value": "456 GB/s" }, { "label": "Memory bus", "value": "192-bit" }, { "label": "Board power", "value": "200W" }, { "label": "Interface", "value": "PCIe 5.0 x8" }, { "label": "Software", "value": "vLLM (Intel), Vulkan, OpenVINO" } ], "goodFor": [ "Low-cost new 24GB", "Multi-card capacity builds", "14–30B models" ], "notFor": [ "Beginners wanting one-click setup", "Speed-first users" ], "sources": [ "https://www.intel.com/content/www/us/en/products/sku/243916/intel-arc-pro-b60-graphics/specifications.html" ] }, { "slug": "apple-mac-mini-m5-pro", "name": "Mac mini (M5 Pro)", "maker": "Apple", "kind": "Desktop", "page": "https://thecontinuitylayer.com/tools/local-hardware/apple-mac-mini-m5-pro", "url": "https://www.apple.com/mac-mini/", "source": "https://www.apple.com/newsroom/2026/08/apple-unveils-a-more-powerful-mac-mini-featuring-the-all-new-m6-and-m5-pro/", "price": "From $1,699 (24GB; 48GB and 64GB cost more)", "price_band": "From $1,699", "price_as_of": "2026-11-02", "capability": "Up to ~32B models at 4-bit with 64GB; 70B too tight", "forWho": "Mac users who want a small, silent box for mid-size local models", "best_for": "Small, silent, mid-size models", "verdict": "The quietest way to run 27–32B models if you order it with 64GB; it cannot be upgraded later.", "description": [ "The Mac mini with M5 Pro was announced in August 2026 alongside a cheaper M6 model. Its memory is shared between the processor and graphics, so most of it can hold a model.", "Buy at least 48GB, and ideally 64GB, for local AI: that runs 27–32B models well and some mixture-of-experts models a little larger. It is quiet and uses little power. Memory cannot be upgraded later, and 64GB is only offered on some chip options." ], "specs": [ { "label": "Chip", "value": "Apple M5 Pro" }, { "label": "Memory", "value": "24GB, 48GB or 64GB unified" }, { "label": "Memory bandwidth", "value": "307 GB/s" }, { "label": "Storage", "value": "512GB to 8TB" }, { "label": "Released", "value": "September 2026" }, { "label": "Software", "value": "MLX, LM Studio, Ollama, llama.cpp" } ], "goodFor": [ "Quiet desk setups", "27–32B models at 64GB", "Low power use" ], "notFor": [ "70B-plus models", "Upgrading memory later" ], "sources": [ "https://www.apple.com/newsroom/2026/08/apple-unveils-a-more-powerful-mac-mini-featuring-the-all-new-m6-and-m5-pro/", "https://www.apple.com/mac-mini/specs/" ] }, { "slug": "apple-macbook-pro-m5-max", "name": "MacBook Pro (M5 Max)", "maker": "Apple", "kind": "Laptop", "page": "https://thecontinuitylayer.com/tools/local-hardware/apple-macbook-pro-m5-max", "url": "https://www.apple.com/macbook-pro/", "source": "https://www.apple.com/newsroom/2026/03/apple-introduces-macbook-pro-with-all-new-m5-pro-and-m5-max/", "price": "From $3,599 (14-inch, 36GB); 128GB costs more", "price_band": "From $3,599 (14-inch) or $3,899 (16-inch)", "price_as_of": "2026-11-02", "capability": "With 128GB: ~120B mixture-of-experts models and 70B at 4-bit", "forWho": "People who need large local models on a laptop they can carry", "best_for": "Large models on a laptop", "verdict": "The only laptop here that runs 70B-class models; you pay for the portability.", "description": [ "The MacBook Pro with M5 Max launched in March 2026 in 14-inch and 16-inch sizes. It is the only laptop here that can be configured with 128GB of unified memory, which requires the 40-core GPU version.", "At 128GB it runs gpt-oss-120b and 70B models at 4-bit, which no other laptop does well. The 36GB base model is limited to about 30B. It gets warm and loud under long AI runs, and the 128GB configurations are expensive. Apple also announced pricier MacBook Pro models with a PROLED XDR display in October 2026, starting at $4,499 with M5 Max." ], "specs": [ { "label": "Chip", "value": "Apple M5 Max" }, { "label": "Memory", "value": "36GB to 128GB unified" }, { "label": "Sizes", "value": "14-inch and 16-inch" }, { "label": "Released", "value": "March 2026" }, { "label": "Software", "value": "MLX, LM Studio, Ollama, llama.cpp" } ], "goodFor": [ "Large models on the go", "Private work while traveling", "One machine for everything" ], "notFor": [ "Budget buyers", "Long unattended runs on battery" ], "sources": [ "https://www.apple.com/newsroom/2026/03/apple-introduces-macbook-pro-with-all-new-m5-pro-and-m5-max/", "https://www.apple.com/macbook-pro/specs/", "https://www.apple.com/newsroom/2026/10/apple-unveils-groundbreaking-macbook-pro-with-proled-xdr-display/" ] }, { "slug": "apple-mac-studio-m5-max", "name": "Mac Studio (M5 Max)", "maker": "Apple", "kind": "Desktop", "page": "https://thecontinuitylayer.com/tools/local-hardware/apple-mac-studio-m5-max", "url": "https://www.apple.com/mac-studio/", "source": "https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/", "price": "From $2,499; 128GB configurations cost more", "price_band": "From $2,499 (36GB); 128GB costs more", "price_as_of": "2026-11-02", "capability": "With 128GB: gpt-oss-120b and 70B models at 4-bit", "forWho": "People who want the fastest, quietest 128GB box for local models", "best_for": "Fastest quiet 128GB box", "verdict": "The best all-round 128GB desk machine for 70B–120B models, faster and quieter than the mini PCs.", "description": [ "The Mac Studio with M5 Max went on sale in September 2026. It supports up to 128GB of unified memory with up to 614 GB/s of bandwidth on the 40-core GPU version, more than double the DGX Spark.", "That bandwidth makes it the fastest of the 128GB machines at generating text, and it is near silent. It costs more than Strix Halo mini PCs at the same memory size, and memory is fixed at purchase." ], "specs": [ { "label": "Chip", "value": "Apple M5 Max" }, { "label": "Memory", "value": "Up to 128GB unified" }, { "label": "Memory bandwidth", "value": "Up to 614 GB/s (460 GB/s on 32-core GPU)" }, { "label": "Released", "value": "September 2026" }, { "label": "Software", "value": "MLX, LM Studio, Ollama, llama.cpp" } ], "goodFor": [ "Fast 70B–120B models", "Quiet always-on desk machine", "Mac software users" ], "notFor": [ "Lowest cost per GB", "CUDA-only tools" ], "sources": [ "https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/", "https://www.apple.com/mac-studio/specs/", "https://www.apple.com/newsroom/2026/09/the-new-mac-mini-and-mac-studio-are-available-today/" ] }, { "slug": "apple-mac-studio-m5-ultra", "name": "Mac Studio (M5 Ultra)", "maker": "Apple", "kind": "Desktop", "page": "https://thecontinuitylayer.com/tools/local-hardware/apple-mac-studio-m5-ultra", "url": "https://www.apple.com/mac-studio/", "source": "https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/", "price": "From $5,499 (96GB); larger memory costs more", "price_band": "From $5,499 (96GB)", "price_as_of": "2026-11-02", "capability": "With 256–512GB: 200B–400B-plus models at 4-bit", "forWho": "Professionals who want the largest open models on one desk machine", "best_for": "Largest models on one desk", "verdict": "The one machine here that runs 200B-plus models on its own, and far more than most home users need.", "description": [ "The M5 Ultra Mac Studio is Apple's largest desktop chip, starting at 96GB and configurable to 512GB of unified memory, with up to 1.2 TB/s of memory bandwidth.", "It is one of very few single machines that can hold the biggest open models without a server rack. Prompt processing on very long inputs is slower than on NVIDIA GPUs. Apple said the 512GB option would arrive in late October 2026." ], "specs": [ { "label": "Chip", "value": "Apple M5 Ultra" }, { "label": "Memory", "value": "96GB to 512GB unified" }, { "label": "Memory bandwidth", "value": "Up to 1.2 TB/s" }, { "label": "Released", "value": "September 2026" } ], "goodFor": [ "200B-plus models on one machine", "Running several models at once", "Quiet operation" ], "notFor": [ "Most home users", "Anyone under a $5,000 budget" ], "sources": [ "https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/", "https://www.apple.com/mac-studio/", "https://www.apple.com/newsroom/2026/09/the-new-mac-mini-and-mac-studio-are-available-today/" ] }, { "slug": "framework-desktop", "name": "Framework Desktop", "maker": "Framework", "kind": "Mini PC", "page": "https://thecontinuitylayer.com/tools/local-hardware/framework-desktop", "url": "https://frame.work/desktop", "source": "https://frame.work/products/framework-desktop-mainboard-amd-ryzen-ai-max-300-series", "price": "$3,149 (128GB mainboard only; full system costs more)", "price_band": "", "price_as_of": "", "capability": "With 128GB: gpt-oss-120b and 70B models at 4-bit, slowly for 70B", "forWho": "Tinkerers who want a 128GB Strix Halo machine from a repair-friendly maker", "best_for": "Cheapest 128GB, repair-friendly", "verdict": "The low-cost 128GB route for Linux tinkerers, strongest on mixture-of-experts models.", "description": [ "The Framework Desktop is a small PC built around AMD's Ryzen AI Max+ 395 (Strix Halo) chip with up to 128GB of shared memory. Framework sells it as a kit or as a bare mainboard.", "Strix Halo machines are a low-cost way to get 128GB for local AI, and mixture-of-experts models like gpt-oss-120b run well. Dense 70B models are slow. Framework raised prices in 2026 because of rising memory costs." ], "specs": [ { "label": "Chip", "value": "AMD Ryzen AI Max+ 395" }, { "label": "Memory", "value": "Up to 128GB LPDDR5x-8000 (soldered)" }, { "label": "Software", "value": "ROCm, Vulkan; Linux or Windows" } ], "goodFor": [ "Low-cost 128GB route", "Mixture-of-experts models", "Linux users" ], "notFor": [ "Fast dense 70B", "People who want plug-and-play" ], "sources": [ "https://frame.work/products/framework-desktop-mainboard-amd-ryzen-ai-max-300-series", "https://frame.work/desktop?tab=specs" ] }, { "slug": "gmktec-evo-x2", "name": "GMKtec EVO-X2", "maker": "GMKtec", "kind": "Mini PC", "page": "https://thecontinuitylayer.com/tools/local-hardware/gmktec-evo-x2", "url": "https://www.gmktec.com/products/amd-ryzen%e2%84%a2-ai-max-395-evo-x2-ai-mini-pc", "source": "https://www.gmktec.com/products/amd-ryzen%e2%84%a2-ai-max-395-evo-x2-ai-mini-pc", "price": "$3,499.99–$3,649.99 (128GB, GMKtec US store)", "price_band": "", "price_as_of": "", "capability": "With 128GB: gpt-oss-120b and 70B models at 4-bit, slowly for 70B", "forWho": "Buyers who want a ready-made 128GB Strix Halo box with Windows installed", "best_for": "Ready-made 128GB with Windows", "verdict": "A plug-in 128GB Strix Halo box; good for mixture-of-experts models, slow on dense 70B.", "description": [ "The EVO-X2 is a compact Ryzen AI Max+ 395 mini PC sold with 64GB or 128GB of memory and Windows 11 Pro. It was one of the first Strix Halo machines on sale.", "It shares the same chip as the Framework Desktop, so model speed is similar. Prices change often with promotions at GMKtec's own store, and the 64GB version starts lower." ], "specs": [ { "label": "Chip", "value": "AMD Ryzen AI Max+ 395" }, { "label": "Memory", "value": "64GB or 128GB LPDDR5X-8000 (soldered)" }, { "label": "Storage", "value": "Two M.2 2280 slots (PCIe 4.0)" }, { "label": "Networking", "value": "2.5GbE, Wi-Fi 7" } ], "goodFor": [ "Ready-to-use 128GB box", "Mixture-of-experts models", "Windows users" ], "notFor": [ "Fast dense 70B", "Buyers who want a long warranty" ], "sources": [ "https://www.gmktec.com/products/amd-ryzen%e2%84%a2-ai-max-395-evo-x2-ai-mini-pc" ] }, { "slug": "minisforum-ms-s1-max", "name": "MINISFORUM MS-S1 MAX", "maker": "MINISFORUM", "kind": "Mini PC", "page": "https://thecontinuitylayer.com/tools/local-hardware/minisforum-ms-s1-max", "url": "https://store.minisforum.com/products/minisforum-ms-s1-max-mini-pc", "source": "https://store.minisforum.com/products/minisforum-ms-s1-max-mini-pc", "price": "varies (Minisforum's own listings show different prices)", "price_band": "$3,919–$4,199 (128GB, sale prices shown; regular $4,899–$5,249)", "price_as_of": "2026-11-02", "capability": "With 128GB: gpt-oss-120b and 70B models at 4-bit, slowly for 70B", "forWho": "Small-office users who want a Strix Halo workstation with fast networking", "best_for": "128GB workstation, fast networking", "verdict": "The 128GB mini PC to pick if you want expansion and fast networking, with prices that move often.", "description": [ "The MS-S1 MAX is a workstation-style mini PC with the Ryzen AI Max+ 395 and 128GB of memory. It adds two 80 Gbps USB4 v2 ports, useful for linking two units.", "Model speed matches other Strix Halo machines. Minisforum's store prices change often with sales. A 64GB version exists but limits you to about 30–70B at tight settings, and its memory cannot be upgraded." ], "specs": [ { "label": "Chip", "value": "AMD Ryzen AI Max+ 395" }, { "label": "Memory", "value": "128GB LPDDR5x-8000 (64GB option)" }, { "label": "Storage", "value": "2TB SSD" }, { "label": "Ports", "value": "2x USB4 v2 (80 Gbps)" }, { "label": "Software", "value": "ROCm, Vulkan; Linux or Windows" } ], "goodFor": [ "128GB with good expansion", "Linking two boxes", "Mixture-of-experts models" ], "notFor": [ "Fast dense 70B", "Buyers who want a fixed, predictable price" ], "sources": [ "https://store.minisforum.com/products/minisforum-ms-s1-max-mini-pc", "https://www.minisforum.com/products/ms-s1-max" ] }, { "slug": "nvidia-dgx-spark", "name": "NVIDIA DGX Spark", "maker": "NVIDIA", "kind": "Mini PC", "page": "https://thecontinuitylayer.com/tools/local-hardware/nvidia-dgx-spark", "url": "https://www.nvidia.com/en-us/products/workstations/dgx-spark/", "source": "https://blogs.nvidia.com/blog/local-ai-dgx-spark-64gb-sync/", "price": "$4,999 (64GB) or $6,950 (128GB)", "price_band": "From $4,999 (64GB)", "price_as_of": "2026-11-02", "capability": "With 128GB: gpt-oss-120b and 70B models at 4-bit", "forWho": "Developers who need NVIDIA's CUDA software stack in a desktop box", "best_for": "CUDA development on a desk", "verdict": "Buy it for the NVIDIA software stack, not for value: the 128GB mini PCs cost far less.", "description": [ "The DGX Spark is a small desktop built on NVIDIA's GB10 chip with 128GB of shared memory, also sold under partner brands such as ASUS, Dell and HP. It runs NVIDIA's Linux software stack with full CUDA support.", "Its strengths are CUDA compatibility and fast prompt processing, not raw generation speed. NVIDIA's marketplace lists the 128GB model at $6,950, up from $3,999 at launch, and a 64GB model starting at $4,999 went on sale from partners on 23 October 2026." ], "specs": [ { "label": "Chip", "value": "NVIDIA GB10 Grace Blackwell" }, { "label": "Memory", "value": "128GB LPDDR5x unified (64GB model available)" }, { "label": "Memory bandwidth", "value": "273 GB/s" }, { "label": "AI performance", "value": "Up to 1 petaflop (FP4)" }, { "label": "Software", "value": "DGX OS (Linux), CUDA" } ], "goodFor": [ "CUDA development on a desk", "Fast prompt processing", "Fine-tuning experiments" ], "notFor": [ "Value buyers (Strix Halo is far cheaper)", "Windows or gaming use" ], "sources": [ "https://blogs.nvidia.com/blog/local-ai-dgx-spark-64gb-sync/", "https://marketplace.nvidia.com/en-us/enterprise/personal-ai-supercomputers/dgx-spark/", "https://www.nvidia.com/en-us/products/workstations/dgx-spark/", "https://docs.nvidia.com/dgx/dgx-spark/hardware.html" ] } ] }, { "slug": "gpu-hosting", "name": "GPU hosting", "page": "https://thecontinuitylayer.com/tools/gpu-hosting", "description": "GPU clouds that rent graphics cards by the hour for running open AI models, compared in one table: published price band, what each is for, and our one-line verdict.", "lastReviewed": "2026-10-31", "capabilityLabel": "GPUs", "priceLabel": "Price per hour", "entries": [ { "slug": "runpod", "name": "Runpod", "maker": "Runpod", "kind": "GPU cloud", "page": "https://thecontinuitylayer.com/tools/gpu-hosting/runpod", "url": "https://www.runpod.io/pricing", "source": "https://www.runpod.io/pricing", "price": "RTX 4090 from $0.34/hr; H100 PCIe from $1.99/hr (Community Cloud)", "price_band": "$0.27–$7.99/hr per GPU", "price_as_of": "2026-11-02", "capability": "Consumer and data-center NVIDIA GPUs, from RTX 4090 to H100 and newer Blackwell cards", "forWho": "Builders who want a cheap GPU by the hour or a serverless endpoint for an open model", "best_for": "Cheap GPU by the hour, serverless", "verdict": "The default place to try an open model for a few dollars, with serverless endpoints when you outgrow a pod.", "description": [ "Runpod rents GPUs in three main forms: Pods (a GPU machine you control), Serverless endpoints that scale up and down with traffic, and multi-node clusters. Prices are shown per hour but billed by the millisecond, and the same GPU comes in a cheaper Community Cloud tier and a pricier Secure Cloud tier run in vetted data centers.", "For open models, most people start a Pod from a ready-made template for vLLM, Ollama or a text-generation web UI, or deploy a model to a Serverless endpoint. Prices change often and depend on availability, so check the live page. Community Cloud hosts vary in reliability, so use Secure Cloud for anything customer-facing." ], "specs": [ { "label": "Billing", "value": "Hourly rates, billed by the millisecond" }, { "label": "RTX 4090", "value": "$0.34/hr Community, $0.89/hr Secure" }, { "label": "H100 PCIe", "value": "$1.99/hr Community, $2.89/hr Secure" }, { "label": "H100 SXM", "value": "$2.69/hr Community, $3.99/hr Secure" }, { "label": "Products", "value": "Pods, Serverless, Instant Clusters" }, { "label": "Tiers", "value": "Community Cloud and Secure Cloud" } ], "goodFor": [ "Trying an open model for a few dollars", "Serverless inference that scales to zero", "Fine-tuning runs on one or a few GPUs" ], "notFor": [ "Teams needing fixed long-term prices", "Workloads that cannot tolerate a Community host going offline" ], "sources": [ "https://www.runpod.io/pricing", "https://www.runpod.io/gpu-models/rtx-4090", "https://www.runpod.io/gpu-models/h100-pcie" ] }, { "slug": "vast-ai", "name": "Vast.ai", "maker": "Vast.ai", "kind": "Marketplace", "page": "https://thecontinuitylayer.com/tools/gpu-hosting/vast-ai", "url": "https://vast.ai/pricing", "source": "https://vast.ai/pricing/gpu/RTX-4090", "price": "Varies; marketplace prices for each GPU change with supply and demand", "price_band": "", "price_as_of": "", "capability": "Marketplace of consumer and data-center GPUs from independent hosts, from older RTX cards to H100", "forWho": "Cost-conscious tinkerers and researchers who are comfortable picking a host themselves", "best_for": "Lowest marketplace prices", "verdict": "Usually the cheapest hour for a 24GB card, if you accept picking a host and the odd interruption.", "description": [ "Vast.ai is a marketplace where independent hosts and data centers list GPU machines and you rent them. You can rent on demand, bid for cheaper interruptible instances, or reserve a machine for a discount.", "It is often among the cheapest places to run an open model on a rented GPU, but quality varies from listing to listing. Check each host's reliability score and verified status, and expect to set up your own container. Prices move with supply and demand, so check the live listings for current rates." ], "specs": [ { "label": "Rental types", "value": "On-demand, interruptible (spot), reserved" }, { "label": "RTX 4090", "value": "Varies (marketplace)" }, { "label": "H100 SXM", "value": "Varies (marketplace)" }, { "label": "H100 PCIe", "value": "Varies (marketplace)" } ], "goodFor": [ "Low hourly prices for a 24GB card", "Batch jobs that can restart if interrupted", "Experimenting with many GPU types" ], "notFor": [ "Production endpoints needing steady uptime", "Sensitive data on unverified hosts" ], "sources": [ "https://vast.ai/pricing/gpu/RTX-4090", "https://vast.ai/products/gpu-cloud", "https://vast.ai/developers/cli" ] }, { "slug": "lambda", "name": "Lambda", "maker": "Lambda", "kind": "GPU cloud", "page": "https://thecontinuitylayer.com/tools/gpu-hosting/lambda", "url": "https://lambda.ai/pricing", "source": "https://lambda.ai/instances", "price": "H100 SXM from $3.99/hr; B200 from $6.69/hr (per GPU, on demand)", "price_band": "$0.79–$6.69/hr per GPU", "price_as_of": "2026-11-02", "capability": "NVIDIA data-center GPUs including B200, H100, A100 and GH200, in 1 to 8 GPU instances and clusters", "forWho": "Teams that want plain, well-run NVIDIA machines for training or serving larger open models", "best_for": "Plain NVIDIA machines for teams", "verdict": "Simple, well-run data-center GPUs for teams; too much machine for hobby budgets.", "description": [ "Lambda is a GPU cloud built for machine-learning work. It rents on-demand instances with 1, 2, 4 or 8 GPUs, plus 1-Click Clusters and reserved capacity for larger jobs. Instances come with a preinstalled ML software stack.", "For open models it is a straightforward place to run vLLM or similar on an H100 or B200 with no egress fees. Prices vary by instance configuration, and popular GPU types can be out of stock in some regions. There is no cheap consumer-card tier." ], "specs": [ { "label": "H100 SXM", "value": "$3.99 per GPU-hour on demand" }, { "label": "B200 SXM6", "value": "$6.69 per GPU-hour on demand" }, { "label": "Instance sizes", "value": "1x, 2x, 4x and 8x GPU" }, { "label": "Billing", "value": "Per minute on instances" }, { "label": "Egress", "value": "No egress fees" }, { "label": "Larger scale", "value": "1-Click Clusters and reserved contracts" } ], "goodFor": [ "Serving 70B-class and larger open models", "Multi-GPU fine-tuning", "Teams wanting a simple ML-focused cloud" ], "notFor": [ "Hobbyists on a small budget", "Anyone wanting serverless scale-to-zero" ], "sources": [ "https://lambda.ai/instances", "https://lambda.ai/pricing" ] }, { "slug": "coreweave", "name": "CoreWeave", "maker": "CoreWeave", "kind": "GPU cloud", "page": "https://thecontinuitylayer.com/tools/gpu-hosting/coreweave", "url": "https://www.coreweave.com/pricing", "source": "https://www.coreweave.com/pricing", "price": "HGX H100 (8 GPUs) $49.24/hr; HGX B200 (8 GPUs) $68.80/hr on demand", "price_band": "$49.24–$68.80/hr per 8-GPU server", "price_as_of": "2026-11-02", "capability": "Large NVIDIA clusters: HGX H100, H200, B200 and GB200 NVL72 systems", "forWho": "Companies running open models at large scale that need whole GPU servers and clusters", "best_for": "Whole servers and clusters", "verdict": "Built for companies renting whole servers and clusters, not for a single-GPU experiment.", "description": [ "CoreWeave is a large specialist GPU cloud that rents full NVIDIA servers and clusters, mainly to AI labs and enterprises. Its pricing is listed per instance, where one instance is usually an 8-GPU server.", "It suits teams that already run Kubernetes and need many GPUs for serving or training. It is not built for someone who wants a single GPU for an afternoon, and the best rates come with long contracts arranged through sales." ], "specs": [ { "label": "HGX H100", "value": "$49.24/hr per 8-GPU instance on demand" }, { "label": "HGX B200", "value": "$68.80/hr per 8-GPU instance on demand" }, { "label": "GB200 NVL72", "value": "$42.00/hr per 4-GPU instance" }, { "label": "Spot", "value": "Offered on some instance types" }, { "label": "Platform", "value": "Kubernetes-based" } ], "goodFor": [ "Large-scale inference fleets", "Multi-node training", "Enterprises with committed spend" ], "notFor": [ "Single-GPU experiments", "Individuals and small hobby projects" ], "sources": [ "https://www.coreweave.com/pricing", "https://docs.coreweave.com/platform/instances/gpu-instances" ] }, { "slug": "together-ai", "name": "Together AI", "maker": "Together AI", "kind": "GPU cloud", "page": "https://thecontinuitylayer.com/tools/gpu-hosting/together-ai", "url": "https://www.together.ai/pricing", "source": "https://www.together.ai/pricing", "price": "GPU clusters: H100 $3.99/hr, B200 $8.19/hr per GPU on demand", "price_band": "$3.99–$8.19/hr per GPU (clusters)", "price_as_of": "2026-11-02", "capability": "NVIDIA H100, H200, B200 and GB200 as dedicated endpoints or self-service clusters", "forWho": "Teams that want a dedicated, managed endpoint for an open model without running servers", "best_for": "Managed endpoint for an open model", "verdict": "The easiest step from a per-token API to your own dedicated capacity, at a premium over raw GPU hours.", "description": [ "Together AI runs open models through a per-token API, and also rents GPUs in two ways: dedicated endpoints, where Together hosts a model you choose on GPUs reserved for you, and GPU clusters you control yourself.", "Dedicated endpoints remove most of the setup work of serving an open model, at a higher hourly price than a bare GPU. Its published per-hour rates differ between the pricing page, docs and announcements, and reserved rates are lower for longer terms, so confirm the current figure before committing." ], "specs": [ { "label": "Cluster H100", "value": "$3.99 per GPU-hour on demand" }, { "label": "Cluster B200", "value": "$8.19 per GPU-hour on demand" }, { "label": "Reserved H100", "value": "From $3.19 per GPU-hour, by term" }, { "label": "Dedicated endpoints", "value": "Billed per GPU-hour per replica" }, { "label": "Other products", "value": "Per-token serverless API, fine-tuning" } ], "goodFor": [ "Managed hosting of an open model", "Moving from a per-token API to dedicated capacity", "Short cluster reservations" ], "notFor": [ "The cheapest possible GPU hour", "Consumer-card workloads" ], "sources": [ "https://www.together.ai/gpu/nvidia-h100", "https://www.together.ai/gpu/nvidia-hgx-b200", "https://www.together.ai/pricing", "https://docs.together.ai/docs/dedicated-endpoints/pricing", "https://www.together.ai/gpu-clusters" ] }, { "slug": "novita-ai", "name": "Novita AI", "maker": "Novita AI", "kind": "GPU cloud", "page": "https://thecontinuitylayer.com/tools/gpu-hosting/novita-ai", "url": "https://novita.ai/pricing", "source": "https://novita.ai/gpu-instance/gpu/rtx-4090", "price": "RTX 4090 $0.67/hr on demand; spot instances about half price", "price_band": "", "price_as_of": "", "capability": "Consumer and data-center GPUs such as RTX 4090, RTX 5090, H100 and H200, plus a per-token model API", "forWho": "Developers who want both an open-model API and rentable GPUs from one account", "best_for": "API plus rentable GPUs, one account", "verdict": "Handy if you want an open-model API and spot GPUs in one place; published rates change often.", "description": [ "Novita AI offers a per-token API for many open-weight models alongside GPU instances, spot instances, serverless GPUs and dedicated model deployments. The GPU side lets you start a container on a chosen card and pay by the hour.", "It is handy if you want to start on the API and move to your own GPU later without changing provider. Its published GPU prices vary with market conditions and differ across its own pages, so check the console for live rates." ], "specs": [ { "label": "RTX 4090", "value": "$0.67/hr on demand (varies with market)" }, { "label": "Spot discount", "value": "Roughly 50% below on demand" }, { "label": "GPU types", "value": "RTX 4090, RTX 5090, H100, H200 and others" }, { "label": "Products", "value": "GPU instances, serverless GPUs, dedicated deployments, model API" } ], "goodFor": [ "Mixing API calls with self-hosted GPUs", "Cheap spot runs on consumer cards" ], "notFor": [ "Buyers who need stable, clearly published rates", "Large multi-node clusters" ], "sources": [ "https://novita.ai/gpu-instance/gpu/rtx-4090", "https://novita.ai/docs/guides/llm-dedicated-endpoint", "https://blogs.novita.ai/spot-vs-on-demand-instances-quick-decision-guide/" ] }, { "slug": "digitalocean-gpu-droplets", "name": "GPU Droplets", "maker": "DigitalOcean", "kind": "GPU cloud", "page": "https://thecontinuitylayer.com/tools/gpu-hosting/digitalocean-gpu-droplets", "url": "https://www.digitalocean.com/pricing/gpu-droplets", "source": "https://docs.digitalocean.com/products/droplets/details/pricing/", "price": "RTX 4000 Ada $0.76/hr; L40S $1.57/hr; H100 $4.41/hr per GPU on demand", "price_band": "$0.76–$4.47/hr per GPU", "price_as_of": "2026-11-02", "capability": "NVIDIA RTX 4000 Ada, L40S, H100, H200 and AMD MI300X", "forWho": "Developers already on DigitalOcean who want a GPU server next to their app", "best_for": "GPU next to an existing app", "verdict": "The sensible pick if you already run on DigitalOcean; not the cheapest H100 hour.", "description": [ "GPU Droplets are DigitalOcean's GPU virtual machines, offered in single-GPU and 8-GPU sizes with both NVIDIA and AMD cards. They sit in the same console as DigitalOcean's regular servers, databases and storage. Paperspace, which DigitalOcean owns, now runs on the same platform.", "For open models, the smaller cards suit 7B to 30B-class models and the H100, H200 and MI300X handle larger ones. Setup is simple, but the big GPUs cost more per hour than at specialist GPU clouds, and 12-month reservations are needed for the lower rates." ], "specs": [ { "label": "RTX 4000 Ada", "value": "$0.76 per GPU-hour" }, { "label": "L40S", "value": "$1.57 per GPU-hour" }, { "label": "H100", "value": "$4.41 per GPU-hour on demand, $3.26 reserved 12 months" }, { "label": "H200", "value": "$4.47 per GPU-hour on demand" }, { "label": "AMD MI300X", "value": "$2.59 per GPU-hour on demand" }, { "label": "Billing", "value": "Per second, with a short minimum charge" } ], "goodFor": [ "Teams already using DigitalOcean", "Mid-size open models on a single card", "Trying AMD MI300X" ], "notFor": [ "The lowest H100 price", "Per-request serverless billing" ], "sources": [ "https://www.digitalocean.com/pricing/gpu-droplets", "https://docs.digitalocean.com/products/droplets/details/pricing/", "https://www.digitalocean.com/products/gpu-droplets" ] }, { "slug": "hyperbolic", "name": "Hyperbolic", "maker": "Hyperbolic", "kind": "Marketplace", "page": "https://thecontinuitylayer.com/tools/gpu-hosting/hyperbolic", "url": "https://www.hyperbolic.ai/marketplace", "source": "https://www.hyperbolic.ai/marketplace", "price": "H100 SXM from $3.19/hr; H200 from $3.99/hr; B200 from $5.99/hr", "price_band": "", "price_as_of": "", "capability": "Rentable H100, H200 and B200 GPUs from a supplier marketplace, plus an open-model API", "forWho": "Developers who want on-demand data-center GPUs without a contract", "best_for": "Short H100 and H200 rentals", "verdict": "On-demand data-center GPUs without a contract, aimed at developers rather than consumer-card budgets.", "description": [ "Hyperbolic rents GPUs on demand through a marketplace that pools capacity from several suppliers, and also runs a per-token API for open models. You pick a GPU type, launch a machine and pay by the hour.", "Its listed starting prices refresh weekly from the best supplier rates, and are well above the lower figures in its older blog posts, so check the live marketplace. It is useful for short runs on H100 or H200 cards, less so for long-term reserved capacity." ], "specs": [ { "label": "H100 SXM", "value": "From $3.19/hr" }, { "label": "H200", "value": "From $3.99/hr" }, { "label": "B200", "value": "From $5.99/hr" }, { "label": "Pricing", "value": "Refreshed weekly from supplier rates" }, { "label": "Other products", "value": "Per-token inference API for open models" } ], "goodFor": [ "Short H100 or H200 rentals", "Comparing an API and a rented GPU from one provider" ], "notFor": [ "Consumer-card budgets", "Fixed long-term pricing" ], "sources": [ "https://www.hyperbolic.ai/marketplace", "https://www.hyperbolic.ai/blog/introducing-hyperbolic-on-demand-cloud" ] }, { "slug": "modal", "name": "Modal", "maker": "Modal Labs", "kind": "Serverless GPU", "page": "https://thecontinuitylayer.com/tools/gpu-hosting/modal", "url": "https://modal.com/pricing", "source": "https://modal.com/pricing", "price": "L4 $0.000222/sec (about $0.80/hr); A100 80GB $0.000694/sec (about $2.50/hr); H100 $0.001097/sec (about $3.95/hr)", "price_band": "$0.000164–$0.001736 per GPU-second (about $0.59–$6.25/hr, converted)", "price_as_of": "2026-11-02", "capability": "Serverless NVIDIA GPUs: T4, L4, A10, L40S, A100, H100, H200, B200 and B300", "forWho": "Python developers who want to serve an open model as code, paying only while it runs", "best_for": "Serve a model as code, pay while it runs", "verdict": "Best for spiky or low-traffic endpoints written in Python; a rented GPU is cheaper for steady load.", "description": [ "Modal is a serverless platform where you define a function or web endpoint in Python, attach a GPU type, and Modal runs it in the cloud. It starts containers on demand and scales them back to zero when idle, billing per second of use.", "For open models, you typically wrap vLLM or a similar server in a Modal app and deploy it as an endpoint. Cold starts and keeping containers warm are the main things to plan for, since warm idle time is billed. Choosing a specific region adds a price multiplier." ], "specs": [ { "label": "Billing", "value": "Per second, scales to zero" }, { "label": "L4", "value": "$0.000222/sec (about $0.80/hr)" }, { "label": "L40S", "value": "$0.000542/sec (about $1.95/hr)" }, { "label": "A100 80GB", "value": "$0.000694/sec (about $2.50/hr)" }, { "label": "H100", "value": "$0.001097/sec (about $3.95/hr)" }, { "label": "B200", "value": "$0.001736/sec (about $6.25/hr)" }, { "label": "Free credits", "value": "$30 of free compute per month on the Starter plan" } ], "goodFor": [ "Spiky or low-traffic inference endpoints", "Developers who prefer code over consoles", "Batch jobs that fan out across many GPUs" ], "notFor": [ "Steady 24/7 load, where a rented GPU is cheaper", "People who want a plain SSH machine" ], "sources": [ "https://modal.com/pricing", "https://modal.com/products/notebooks" ] }, { "slug": "nebius", "name": "Nebius AI Cloud", "maker": "Nebius", "kind": "GPU cloud", "page": "https://thecontinuitylayer.com/tools/gpu-hosting/nebius", "url": "https://nebius.com/prices", "source": "https://nebius.com/prices", "price": "H100 $4.50/hr; H200 $5.40/hr; B200 $8.50/hr per GPU on demand", "price_band": "$1.55–$8.50/hr per GPU", "price_as_of": "2026-11-02", "capability": "NVIDIA H100, H200, B200 and B300 with NVLink, from single GPUs to large clusters", "forWho": "Teams that want a full-featured AI cloud with self-service access to recent NVIDIA GPUs", "best_for": "Recent NVIDIA GPUs, self-service", "verdict": "A full AI cloud with self-service H200 and B200 access, better for sustained work than a quick test.", "description": [ "Nebius is an AI-focused cloud that rents NVIDIA data-center GPUs as virtual machines and managed Kubernetes clusters, with storage and networking around them. It also offers preemptible GPUs at much lower prices.", "It suits teams serving or fine-tuning larger open models who want more of a full cloud than a bare GPU rental. Preemptible prices move with capacity and demand, and preemptible machines can be stopped at short notice, so they suit batch work more than live endpoints." ], "specs": [ { "label": "H100 NVLink", "value": "$4.50 per GPU-hour on demand" }, { "label": "H200 NVLink", "value": "$5.40 per GPU-hour on demand" }, { "label": "B200 NVLink", "value": "$8.50 per GPU-hour on demand" }, { "label": "B300", "value": "$9.50 per GPU-hour on demand" }, { "label": "Preemptible", "value": "From $0.79 per GPU-hour (H100, H200); varies with demand" }, { "label": "Platform", "value": "VMs, managed Kubernetes, storage" } ], "goodFor": [ "Larger open models needing H200 or B200", "Multi-GPU fine-tuning", "Cheap preemptible batch jobs" ], "notFor": [ "Consumer-card budgets", "Single short experiments where marketplaces are cheaper" ], "sources": [ "https://docs.nebius.com/compute/resources/pricing", "https://nebius.com/prices", "https://nebius.com/preemptible-vms" ] } ] }, { "slug": "routers", "name": "Routers", "page": "https://thecontinuitylayer.com/tools/routers", "description": "AI model routers and gateways compared in one table: one API key for many models, the fee each charges, what each is for, and our one-line verdict.", "lastReviewed": "2026-10-31", "capabilityLabel": "Routes to", "priceLabel": "Fees", "entries": [ { "slug": "openrouter", "name": "OpenRouter", "maker": "OpenRouter", "kind": "Router", "page": "https://thecontinuitylayer.com/tools/routers/openrouter", "url": "https://openrouter.ai/pricing", "source": "https://openrouter.ai/blog/announcements/simplifying-our-platform-fee/", "price": "5.5% fee on credit purchases; model prices passed through at provider rates", "price_band": "5.5% fee on credit purchases (5% by crypto); no markup stated", "price_as_of": "2026-11-02", "capability": "600+ models from 80+ providers behind one OpenAI-compatible API (vendor figures)", "forWho": "Anyone who wants one key and one bill for most popular and open models", "best_for": "One key for most models", "verdict": "The baseline: the widest model list and the simplest bill, at a 5.5% fee on every top-up.", "description": [ "OpenRouter is a hosted service that gives you one API key and one prepaid balance for hundreds of models from many different providers. You send requests in the OpenAI format and choose a model by name, and OpenRouter forwards the request to a provider that serves it. It is the reference point for the other tools on this page.", "For open models it is an easy way to try many options without signing up with each host. Free models are limited to 50 requests a day, and buying $10 of credits raises that to 1,000 a day. The 5.5% fee (8% on the Business plan) is charged when you buy credits, not per request. Bring-your-own-key traffic is free up to $25,000 of list-price inference a month, then a 5% fee applies." ], "specs": [ { "label": "Fee", "value": "5.5% on credit purchases (8% on Business)" }, { "label": "Model prices", "value": "Passed through at provider rates" }, { "label": "Free models", "value": "50 requests/day; 1,000/day after buying $10 of credits" }, { "label": "Bring your own key", "value": "Free up to $25,000 of list-price inference a month, then 5%" }, { "label": "API format", "value": "OpenAI-compatible" }, { "label": "Hosting", "value": "Hosted only" } ], "goodFor": [ "Trying many open models with one key", "Free-model experiments", "One prepaid bill across providers" ], "notFor": [ "Teams that must self-host the gateway", "Heavy use where a 5.5% fee adds up" ], "sources": [ "https://openrouter.ai/blog/announcements/simplifying-our-platform-fee/", "https://openrouter.ai/pricing", "https://openrouter.ai/docs/faq", "https://openrouter.ai/about" ] }, { "slug": "requesty", "name": "Requesty", "maker": "Requesty", "kind": "Router", "page": "https://thecontinuitylayer.com/tools/routers/requesty", "url": "https://www.requesty.ai/pricing", "source": "https://www.requesty.ai/pricing", "price": "5% added to model cost on pay-as-you-go", "price_band": "5% markup on provider rates", "price_as_of": "2026-11-02", "capability": "600+ models from many providers through one OpenAI-compatible endpoint (vendor figure)", "forWho": "Builders who want an OpenRouter-style router with routing and spend controls", "best_for": "OpenRouter-style with spend controls", "verdict": "A near drop-in for OpenRouter with routing and spend controls, at a similar cut.", "description": [ "Requesty is a hosted LLM router: one API key and one OpenAI-compatible endpoint that sends your requests to many model providers. It adds features for teams such as fallbacks between providers and spend tracking.", "Pricing works much like OpenRouter's, with 5% added to the provider's per-token price instead of 5.5% on credit purchases, so a model listed at $10 per million tokens costs $10.50. There is a free tier of 200 requests a day on free models with no credit card, and traffic that uses your own provider keys has no markup." ], "specs": [ { "label": "Fee", "value": "5% on model cost" }, { "label": "Free tier", "value": "200 requests/day on free models" }, { "label": "Models", "value": "600+ (vendor figure)" }, { "label": "Bring your own key", "value": "0% markup" }, { "label": "API format", "value": "OpenAI-compatible" }, { "label": "Hosting", "value": "Hosted only" } ], "goodFor": [ "Swapping in for OpenRouter with a similar fee", "Teams wanting fallbacks and spend tracking" ], "notFor": [ "Self-hosting", "Anyone who wants zero markup" ], "sources": [ "https://www.requesty.ai/pricing", "https://www.requesty.ai/models" ] }, { "slug": "kilo-gateway", "name": "Kilo Gateway", "maker": "Kilo Code", "kind": "Router", "page": "https://thecontinuitylayer.com/tools/routers/kilo-gateway", "url": "https://kilo.ai/gateway", "source": "https://kilo.ai/gateway", "price": "No markup on model prices; 5% processing fee on card credit purchases; optional Kilo Pass at $19, $49 or $199/month", "price_band": "No markup on model prices; Kilo Pass from $19/month", "price_as_of": "2026-11-02", "capability": "500+ models from 60+ providers on one pay-as-you-go bill (vendor figure)", "forWho": "Developers, especially Kilo Code users, who want provider prices with no markup", "best_for": "Provider prices, coding agents", "verdict": "Provider list prices with no markup, best if you already use Kilo Code.", "description": [ "Kilo Gateway is a hosted inference API from the company behind the open-source Kilo Code coding agent. It routes to many models through one key and one bill, and says it charges provider list prices with no markup on tokens, including for bring-your-own-key traffic.", "It works as a general router, but it is built around Kilo's coding tools. The optional Kilo Pass subscription turns your payment into credits one for one and adds bonus credits that expire at the end of the month. Kilo charges a 5% processing fee on credits bought by card, including Kilo Pass." ], "specs": [ { "label": "Token markup", "value": "None (vendor states)" }, { "label": "Processing fee", "value": "5% on card purchases of credits" }, { "label": "Bring your own key", "value": "No Kilo markup" }, { "label": "Subscription", "value": "Optional Kilo Pass at $19, $49 or $199/month, with bonus credits" }, { "label": "Models", "value": "500+ from 60+ providers (vendor figure)" }, { "label": "Hosting", "value": "Hosted only" } ], "goodFor": [ "Coding-agent users", "Paying provider list prices", "Bring-your-own-key without a markup" ], "notFor": [ "Teams that must self-host", "Users wanting a gateway independent of a coding tool" ], "sources": [ "https://kilo.ai/inference", "https://kilo.ai/gateway", "https://blog.kilo.ai/p/introducing-kilo-pass" ] }, { "slug": "eden-ai", "name": "Eden AI", "maker": "Eden AI", "kind": "Router", "page": "https://thecontinuitylayer.com/tools/routers/eden-ai", "url": "https://www.edenai.co/pricing", "source": "https://www.edenai.co/pricing", "price": "5.5% platform fee on credit purchases; model prices passed through with no markup", "price_band": "5.5% platform fee on credit purchases; no markup", "price_as_of": "2026-11-02", "capability": "About 1,000 models from 80+ providers, including OCR, translation and audio as well as LLMs (vendor figure)", "forWho": "Builders who need LLMs plus other AI services (OCR, speech, translation) behind one API", "best_for": "LLMs plus OCR, speech, translation", "verdict": "The pick when you need more than text models behind one API; otherwise a plainer router is cheaper.", "description": [ "Eden AI is a hosted gateway with one API to many AI providers. Its chat endpoint follows the OpenAI format, and a separate endpoint covers specialist services such as OCR, translation and audio. There is no subscription, seat fee or minimum commitment on the self-serve plan.", "The fee matches OpenRouter's 5.5% and, according to Eden's pricing docs, is charged at checkout when you buy credits rather than per request. It is broader than a pure LLM router, which helps if your app needs more than text models. Some free models are offered; check the pricing page for current terms." ], "specs": [ { "label": "Fee", "value": "5.5% on credit purchases" }, { "label": "Model prices", "value": "Provider price, no markup" }, { "label": "Models", "value": "About 1,000 from 80+ providers (vendor figure)" }, { "label": "API format", "value": "OpenAI-compatible chat endpoint" }, { "label": "Extras", "value": "OCR, translation, audio via a separate endpoint" }, { "label": "Hosting", "value": "Hosted only" } ], "goodFor": [ "Apps mixing LLMs with OCR or speech", "Pay-as-you-go with no subscription" ], "notFor": [ "Self-hosting", "Users who only need text models and want the lowest fee" ], "sources": [ "https://www.edenai.co/pricing", "https://www.edenai.co/docs/v3/overview/plans-prices", "https://docs.edenai.co/v3/overview/ai-gateway" ] }, { "slug": "litellm", "name": "LiteLLM", "maker": "BerriAI", "kind": "Gateway", "page": "https://thecontinuitylayer.com/tools/routers/litellm", "url": "https://www.litellm.ai/", "source": "https://www.litellm.ai/pricing", "price": "Free and open source (MIT) to self-host; you pay providers directly; enterprise annual license is custom-priced", "price_band": "Free to self-host; enterprise priced by quote", "price_as_of": "2026-11-02", "capability": "100+ LLM providers through one OpenAI-compatible proxy (pricing page; homepage says 140+)", "forWho": "Developers and teams who want to run their own router with no per-request fee", "best_for": "Run your own router, no fee", "verdict": "The standard self-hosted gateway: no per-request fee, but you run and maintain the server.", "description": [ "LiteLLM is an open-source Python library and proxy server that puts many LLM providers behind one OpenAI-style API. You run it yourself, add your own provider keys, and point your apps at it. It works with local servers too, so it can sit in front of models you host on your own hardware.", "The free version includes virtual keys, budgets, rate limits, cost tracking, load balancing, fallbacks and guardrails. You pay nothing to LiteLLM, but you take on running and updating the proxy. Features such as SSO with SCIM, audit logs and secret-manager key rotation need a paid enterprise license, which is an annual fee based on request capacity and support, never per token. Enterprise prices are not published." ], "specs": [ { "label": "License", "value": "MIT (open source)" }, { "label": "Fee", "value": "None for self-hosting" }, { "label": "Providers", "value": "100+" }, { "label": "Included free", "value": "Virtual keys, budgets, rate limits, cost tracking, load balancing, fallbacks, guardrails" }, { "label": "Enterprise", "value": "Annual license, price not published (SSO, SCIM, audit logs, SLA)" }, { "label": "Hosting", "value": "Self-hosted" } ], "goodFor": [ "Self-hosting with no markup", "Mixing cloud APIs with local models", "Per-team budgets" ], "notFor": [ "People who do not want to run a server", "Teams needing a vendor SLA without an enterprise deal" ], "sources": [ "https://www.litellm.ai/pricing", "https://www.litellm.ai/", "https://github.com/BerriAI/litellm" ] }, { "slug": "portkey", "name": "Portkey", "maker": "Portkey", "kind": "Gateway", "page": "https://thecontinuitylayer.com/tools/routers/portkey", "url": "https://portkey.ai/pricing", "source": "https://portkey.ai/pricing", "price": "Open-source gateway free to self-host; hosted plan free for 10,000 logs/month; Production $49/month for 100,000 logs", "price_band": "Free to 10,000 logs/month; $49/month for 100,000", "price_as_of": "2026-11-02", "capability": "Routes to 1,600+ models (GitHub repo figure) across many providers", "forWho": "Teams that want a gateway with logging, guardrails and prompt management", "best_for": "Logging, guardrails, prompts", "verdict": "The gateway to pick for production logging and guardrails; hobby use outgrows the free tier fast.", "description": [ "Portkey is an AI gateway with an open-source core you can self-host for free, plus a hosted service that adds observability, prompt management, guardrails and governance. You bring your own provider keys, and Portkey routes, retries and logs the requests.", "Hosted plans are priced by recorded logs, one per request. The free Developer tier covers 10,000 logs a month with 3-day retention; requests past the limit still go through but are not logged. The Production plan is $49 a month for 100,000 logs, $9 per additional 100,000, with 30-day retention. It is aimed more at production teams than at someone who just wants to try open models." ], "specs": [ { "label": "Open-source gateway", "value": "MIT, free to self-host" }, { "label": "Free hosted tier", "value": "10,000 logs/month, 3-day retention" }, { "label": "Production plan", "value": "$49/month for 100,000 logs, $9 per extra 100,000, 30-day retention" }, { "label": "Models", "value": "1,600+ (GitHub repo figure)" }, { "label": "Keys", "value": "Bring your own provider keys" }, { "label": "Hosting", "value": "Hosted or self-hosted" } ], "goodFor": [ "Production logging and guardrails", "Choice of hosted or self-hosted" ], "notFor": [ "Beginners wanting one prepaid balance", "Hobby use past the free log limit" ], "sources": [ "https://portkey.ai/pricing", "https://portkey.ai/docs/product/observability/logs", "https://github.com/portkey-ai/gateway" ] }, { "slug": "vercel-ai-gateway", "name": "Vercel AI Gateway", "maker": "Vercel", "kind": "Gateway", "page": "https://thecontinuitylayer.com/tools/routers/vercel-ai-gateway", "url": "https://vercel.com/docs/ai-gateway/pricing", "source": "https://vercel.com/docs/ai-gateway/pricing", "price": "No markup or platform fee on tokens, including bring your own key; $5/month free credit on eligible models", "price_band": "No markup or platform fee on tokens", "price_as_of": "2026-11-02", "capability": "Hundreds of models from 20+ providers through one endpoint", "forWho": "Web developers, especially on Vercel, who want provider prices with no markup", "best_for": "Vercel apps, provider prices", "verdict": "No markup and a free tier on some models, strongest if your app already lives on Vercel.", "description": [ "Vercel AI Gateway is a hosted gateway that lets you call models from many providers with one key, billed through your Vercel account. It works with Vercel's AI SDK and also with OpenAI-compatible clients.", "Vercel says it charges provider list prices with no markup, and bring-your-own-key has no gateway fee but needs the paid tier. The free tier is a $5 monthly credit that only covers free-tier-eligible models with lower limits, and it stops once you buy credits." ], "specs": [ { "label": "Token markup", "value": "None" }, { "label": "Bring your own key", "value": "No fee; paid tier only" }, { "label": "Free tier", "value": "$5/month credit, eligible models only" }, { "label": "Models", "value": "Hundreds from 20+ providers (vendor figure)" }, { "label": "Hosting", "value": "Hosted only" } ], "goodFor": [ "Apps already deployed on Vercel", "No-markup access to many models" ], "notFor": [ "Self-hosting", "Generous free use" ], "sources": [ "https://vercel.com/docs/ai-gateway/pricing", "https://vercel.com/ai-gateway", "https://vercel.com/docs/ai-gateway" ] }, { "slug": "cloudflare-ai-gateway", "name": "Cloudflare AI Gateway", "maker": "Cloudflare", "kind": "Gateway", "page": "https://thecontinuitylayer.com/tools/routers/cloudflare-ai-gateway", "url": "https://developers.cloudflare.com/ai-gateway/reference/pricing/", "source": "https://developers.cloudflare.com/ai-gateway/reference/pricing/", "price": "Core features free on all plans; 5% fee on credits bought through Unified Billing; no token markup", "price_band": "Core features free; 5% fee on Unified Billing credits; no markup", "price_as_of": "2026-11-02", "capability": "Proxies 20+ AI providers and Workers AI with caching, rate limits and logs", "forWho": "Developers who already use their own provider keys and want caching, logs and spend limits for free", "best_for": "Caching and logs over your own keys", "verdict": "Free caching, logs and spend limits in front of keys you already hold; not a model catalog.", "description": [ "Cloudflare AI Gateway sits between your app and the AI providers you already use. It adds analytics, caching, rate limiting and logs, and you can keep your own provider keys. Optional Unified Billing lets you prepay Cloudflare and get one bill for several major providers and Workers AI.", "The core gateway is free. For accounts that created their first gateway before September 24, 2026, the free Workers plan stores up to 100,000 logs across all gateways and the paid plan 10,000,000 per gateway. Unified Billing adds 5% to credit purchases, so $100 of credit costs $105. It is less of a model marketplace than OpenRouter: it mainly adds controls to providers you choose yourself." ], "specs": [ { "label": "Core features", "value": "Free on all plans" }, { "label": "Unified Billing fee", "value": "5% on credit purchases" }, { "label": "Free plan logs", "value": "100,000 total across gateways (older accounts)" }, { "label": "Paid plan logs", "value": "10,000,000 per gateway (older accounts)" }, { "label": "Providers", "value": "20+ plus Workers AI" }, { "label": "Hosting", "value": "Hosted only" } ], "goodFor": [ "Free caching and logging over your own keys", "Spend limits and rate limits" ], "notFor": [ "Browsing a large catalog of open models", "Self-hosting" ], "sources": [ "https://developers.cloudflare.com/ai-gateway/reference/pricing/", "https://developers.cloudflare.com/ai-gateway/reference/limits/", "https://developers.cloudflare.com/ai-gateway/usage/providers/" ] } ] } ] }