Open models that run on a 64–128GB computer
thecontinuitylayer.com/open/models/64-128gbChecked 6 November 2026. US dollars. No affiliate links.
22 open models in our catalog fit a 64–128GB computer: Gemma 4, Llama 3, DeepSeek R1, Qwen3.5, Dolphin (community) and 9 more. 15 of them carry a permissive license (Apache 2.0 or MIT), 20 run with one Ollama command, 9 read images. Each figure comes from the maker's page or model card, checked 6 November 2026.
64–128GB: a 64–128gb mac studio or 128gb mini pc. A model that fits a smaller machine fits this one too, so the list runs from the smallest up. Want to filter by license, who set the behavior, or images? Use the model finder.
Gemma 4 E2B
Fits 8GBGoogle DeepMind · 2B effective parameters
Phones, old laptops, quick private chat
- License
- Apache 2.0
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text, Image, Audio
- Context
- 128K
- Download
- 2.9GB at 4-bit (Google's figure)
ollama run gemma4:e2b
Source · checked 6 November 2026 · Gemma 4 graded
Gemma 4 E4B
Fits 8GBGoogle DeepMind · 4B effective parameters
The best first model for an 8GB machine
- License
- Apache 2.0
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text, Image, Audio
- Context
- 128K
- Download
- 4.5GB at 4-bit (Google's figure)
ollama run gemma4:e4b
Source · checked 6 November 2026 · Gemma 4 graded
Llama 3.2 3B
Fits 8GBMeta · 3B parameters
Very old or very small machines
- License
- Llama 3.2 Community License
- License rules
- Acceptable Use Policy; very large services (700M+ monthly users) need Meta's permission
- Behavior
- Maker's own tuning
- Reads
- Text
- Context
- 128K
- Download
- 2.0GB
ollama run llama3.2:3b
Source · checked 6 November 2026
Llama 3.1 8B
Fits 8GBMeta · 8B parameters
The most-downloaded base for community retunes
- License
- Llama 3.1 Community License
- License rules
- Acceptable Use Policy; very large services (700M+ monthly users) need Meta's permission
- Behavior
- Maker's own tuning
- Reads
- Text
- Context
- 128K
- Download
- 4.9GB
ollama run llama3.1:8b
Source · checked 6 November 2026
DeepSeek-R1-0528-Qwen3-8B
Fits 8GBDeepSeek · 8B parameters
Step-by-step reasoning on a small machine (it thinks out loud, so it is slower)
- License
- MIT (built on Qwen3 8B)
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text
- Context
- 128K
- Download
- 5.2GB
ollama run deepseek-r1:8b
Source · checked 6 November 2026
Qwen3.5 9B
Fits 16GBAlibaba (Qwen) · 9B parameters
All-round everyday model with vision and tool use
- License
- Apache 2.0
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text, Image
- Context
- 256K
- Download
- 6.6–7.6GB
- Note
- Loads on 8GB with little room left for long chats; 16GB is comfortable.
ollama run qwen3.5:9b
Source · checked 6 November 2026
Dolphin 3.0 8B
Fits 8GBCognitive Computations · 8B (Llama 3.1 8B base) parameters
A model that follows your system prompt rather than the maker's defaults
- License
- Llama 3.1 Community License (inherited)
- License rules
- Meta's Acceptable Use Policy still applies to a retune of Llama
- Behavior
- Community retune; its page says the system owner sets the guidelines
- Reads
- Text
- Context
- 128K
- Download
- 4.9GB
ollama run dolphin3
Source · checked 6 November 2026
Hermes 3 8B
Fits 8GBNous Research · 8B (Hermes-3-Llama-3.1-8B) parameters
Role-play, fiction and agents that should take direction from you
- License
- Llama 3.1 Community License (inherited)
- License rules
- Meta's Acceptable Use Policy still applies to a retune of Llama
- Behavior
- Community retune; Nous says it aligns the model to the user and gives the user strong steering
- Reads
- Text
- Context
- 128K
- Download
- 4.7GB
ollama run hermes3:8b
Source · checked 6 November 2026
Gemma 4 12B
Fits 16GBGoogle DeepMind · 12B parameters
The sweet spot for a 16GB laptop: documents, images, writing
- License
- Apache 2.0
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text, Image
- Context
- 256K
- Download
- 6.7GB at 4-bit (Google's figure)
ollama run gemma4:12b
Source · checked 6 November 2026 · Gemma 4 graded
gpt-oss-20b
Fits 16GBOpenAI · 21B parameters, 3.6B active
Reasoning and tool use, with adjustable thinking effort
- License
- Apache 2.0
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text
- Context
- 128K
- Download
- 14GB
- Note
- The maker says it runs within 16GB of memory; a 16GB laptop has little left over for other apps.
ollama run gpt-oss:20b
Source · checked 6 November 2026
Phi-4
Fits 16GBMicrosoft · 14B parameters
Logic and short reasoning tasks in English; short context
- License
- MIT
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text
- Context
- 16K
- Download
- 9.1GB
ollama run phi4
Source · checked 6 November 2026
Gemma 4 26B A4B
Fits 24–32GBGoogle DeepMind · 26B parameters, 4B active
Fast on a 24GB card because only 4B parameters work per word
- License
- Apache 2.0
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text, Image
- Context
- 256K
- Download
- 14.4GB at 4-bit (Google's figure)
ollama run gemma4:26b
Source · checked 6 November 2026 · Gemma 4 graded
Gemma 4 31B
Fits 24–32GBGoogle DeepMind · 31B parameters
Best Gemma quality that fits one 24GB card
- License
- Apache 2.0
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text, Image
- Context
- 256K
- Download
- 17.5GB at 4-bit (Google's figure)
ollama run gemma4:31b
Source · checked 6 November 2026 · Gemma 4 graded
Qwen3.8 27B
Fits 24–32GBAlibaba (Qwen) · 27B parameters
The strongest all-rounder for a 24GB card, and coding agents
- License
- Apache 2.0
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text, Image, Video
- Context
- 256K
- Download
- 18GB
ollama run qwen3.8:27b
Source · checked 6 November 2026 · Qwen3.8 graded
Mistral Small 3.2 24B
Fits 24–32GBMistral AI · 24B parameters
European-made all-rounder with reliable function calling
- License
- Apache 2.0
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text, Image
- Context
- 128K
- Download
- 15GB
ollama run mistral-small3.2
Source · checked 6 November 2026
Nemotron 3.5 Lightning
Fits 24–32GBNVIDIA · 30B parameters, 3B active
Long documents on a 32GB-plus graphics card
- License
- OpenMDW-1.1
- License rules
- Keep the agreement and notices; a patent-retaliation clause
- Behavior
- Maker's own tuning
- Reads
- Text
- Context
- Up to 1M (4K by default under 24GB)
- Download
- —
- Note
- NVIDIA lists RTX 5090 and DGX Spark; under 24GB Ollama defaults to a 4K context.
Source · checked 3 November 2026 · Nemotron graded
Olmo 3.1 32B
Fits 24–32GBAi2 · 32B parameters
The fully open choice: data, code and checkpoints are all published
- License
- Apache 2.0 · fully open (data and code too)
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text
- Context
- Not stated on the card
- Download
- —
- Note
- Our estimate: 32B at 4-bit needs about 18–20GB. Ai2 gives no figure.
Source · checked 6 November 2026 · Olmo graded
Devstral Small 2
Fits 24–32GBMistral AI · 24B parameters
Coding agents that explore and edit a whole codebase
- License
- Apache 2.0
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text, Image
- Context
- 384K
- Download
- 15GB
ollama run devstral-small-2
Source · checked 6 November 2026
Qwen3-Coder 30B A3B
Fits 24–32GBAlibaba (Qwen) · 30B parameters, 3.3B active
Fast coding help on a 24GB card: only 3.3B parameters work per word
- License
- Not stated on the Ollama page; read the card
- License rules
- Not read yet: check the Hugging Face card before commercial use
- Behavior
- Maker's own tuning
- Reads
- Text
- Context
- 256K
- Download
- 19GB
ollama run qwen3-coder:30b
Source · checked 6 November 2026
Llama 3.3 70B
Fits 64–128GBMeta · 70B parameters
A large dense model for a 64GB-plus Mac; slow on mini PCs
- License
- Llama 3.3 Community License
- License rules
- Acceptable Use Policy; very large services (700M+ monthly users) need Meta's permission
- Behavior
- Maker's own tuning
- Reads
- Text
- Context
- 128K
- Download
- 43GB
ollama run llama3.3
Source · checked 6 November 2026
gpt-oss-120b
Fits 64–128GBOpenAI · 117B parameters, 5.1B active
Close to a hosted experience on a 128GB machine
- License
- Apache 2.0
- License rules
- None beyond keeping the notices
- Behavior
- Maker's own tuning
- Reads
- Text
- Context
- 128K
- Download
- 65GB
- Note
- The maker targets one 80GB GPU; a 128GB Mac or mini PC runs it.
ollama run gpt-oss:120b
Source · checked 6 November 2026 · gpt-oss graded
Hermes 3 70B
Fits 64–128GBNous Research · 70B (Hermes-3-Llama-3.1-70B) parameters
A large steerable model for a 64GB-plus machine
- License
- Llama 3.1 Community License (inherited)
- License rules
- Meta's Acceptable Use Policy still applies to a retune of Llama
- Behavior
- Community retune; Nous says it aligns the model to the user and gives the user strong steering
- Reads
- Text
- Context
- 128K
- Download
- 40GB
ollama run hermes3:70b
Source · checked 6 November 2026
Other machine sizes
- Fits 8GB · 7 models · most laptops, an 8gb mac, an 8gb graphics card
- Fits 16GB · 11 models · a 16gb mac or laptop, a 12–16gb graphics card
- Fits 24–32GB · 19 models · a 24gb graphics card, a 32gb mac
- Needs a server · 5 models · rented gpus or a host: too big for one home machine
Not sure what your machine has? The will it run page says how to check, and the hardware guide says what to buy, or not.