The Continuity Layer
← The ScorecardCoding agents →Memory systems →Inference APIs →Run your own AI: the finder, the apps, the path →Which open model runs on your computer →Best open model now →
The Scorecard · Open models

Open models: what the license lets you do, what you can download, and what the maker charges

thecontinuitylayer.com/open-models

Qwen, DeepSeek, Mistral, Gemma, gpt-oss, Kimi, GLM, Meta's Muse Spark Open and more, read from their makers' own pages and graded on the same seven questions: the license, whether the weights download, whether the context window is stated, what it takes to run locally, whether the maker publishes a hosted price, whether your prompts train the maker's models, and whether you can fine-tune it and keep the result.

12
Open models graded
7
Tests
84
Sourced answers
Docs
Basis
Read from the makers' own pages on 3 November 2026

Eleven of the twelve download without an access form, ten state a context window, and six carry a plain Apache 2.0 or MIT license: MiMo-V2.6-Pro, Phi-4-reasoning-vision-15B, Gemma 4, DeepSeek-V4.1-Flash, gpt-oss-120b and Olmo 3.1 32B; the other six add conditions, such as a revenue threshold for running the model as a service, or use a permissive license the OSI has not approved. Six makers publish a per-token price for their hosted version, and only two, Xiaomi and Microsoft, say in plain words that prompts to it are not used for training; every grade is read from the maker's own page on the day shown, and no benchmark score appears here.

At a glance

Every product, every test.

✓ yes, ½ partly, ✕ no, – the maker's pages do not say. A dash counts for nothing: a product gets credit only for what its own pages state, so a vague page never outranks a clear one. On the training test the colors flip, because no is the good answer there.

Sort by
Open models: each product's grade on each test, and its score out of all the tests
ProductLicenseDownloadContextRuns locallyHosted priceTrainingFine-tuneScore
MiMo-V2.6-Pro#1 · XiaomiA6½/7
Phi-4-reasoning-vision-15B#2 · MicrosoftA6½/7
Gemma 4#3 · Google DeepMindA−6/7
DeepSeek-V4.1-Flash#4 · DeepSeekB+5½/7
Mistral Large 4#5 · Mistral AIB5/7
gpt-oss-120b#6 · OpenAIB5/7
NVIDIA Nemotron 3.5 Lightning#7 · NVIDIAB5/7
Kimi K3#8 · Moonshot AIB−4½/7
Qwen3.8#9 · Alibaba (Qwen)B−4½/7
GLM-5.3#10 · Z.aiC4/7
Muse Spark Open#11 · MetaC−3½/7
Olmo 3.1 32B#12 · Ai2 (Allen Institute for AI)C−3½/7
Product by product

In score order, best first.

  • MiMo-V2.6-Pro

    #1 · Xiaomi

    The flagship checkpoint of Xiaomi's MiMo-V2.6 series: a sparse mixture-of-experts model with 1.02T total and 42B active parameters that takes text, image, video and audio input, served on the Xiaomi MiMo API platform.

    A6½/7On the 7 tests
    • Is MiMo-V2.6-Pro under an open license?Yes
    • Can you download MiMo-V2.6-Pro's weights?Yes
    • Does the maker state MiMo-V2.6-Pro's context window?Yes
    All 7 answers →

  • Phi-4-reasoning-vision-15B

    #2 · Microsoft

    Microsoft's compact open-weight multimodal reasoning model: a 15B model combining the Phi-4-Reasoning language backbone with a SigLIP-2 image encoder, released 4 March 2026 under MIT.

    A6½/7On the 7 tests
    • Is Phi-4-reasoning-vision-15B under an open license?Yes
    • Can you download Phi-4-reasoning-vision-15B's weights?Yes
    • Does the maker state Phi-4-reasoning-vision-15B's context window?Yes
    All 7 answers →

  • Gemma 4

    #3 · Google DeepMind

    Google DeepMind's family of open models in five sizes (E2B, E4B, 12B, 26B A4B and 31B), with a 256K context window on the medium models and Apache 2.0 licensing; the two largest also run on the Gemini API.

    A−6/7On the 7 tests
    • Is Gemma 4 under an open license?Yes
    • Can you download Gemma 4's weights?Yes
    • Does the maker state Gemma 4's context window?Yes
    All 7 answers →

  • DeepSeek-V4.1-Flash

    #4 · DeepSeek

    DeepSeek's current open release: a multimodal mixture-of-experts model with 552B backbone parameters, served on DeepSeek's API as deepseek-flash (version DeepSeek-V4.1-Flash-1017).

    B+5½/7On the 7 tests
    • Is DeepSeek-V4.1-Flash under an open license?Yes
    • Can you download DeepSeek-V4.1-Flash's weights?Yes
    • Does the maker state DeepSeek-V4.1-Flash's context window?Yes
    All 7 answers →

  • Mistral Large 4

    #5 · Mistral AI

    Mistral AI's 1-trillion-parameter natively multimodal mixture-of-experts model with about 52B active parameters and a 1M context, released on Hugging Face in BF16 and NVFP4; Mistral Small 4 (119B, Apache 2.0) is the smaller open release.

    B5/7On the 7 tests
    • Is Mistral Large 4 under an open license?Partly
    • Can you download Mistral Large 4's weights?Yes
    • Does the maker state Mistral Large 4's context window?Yes
    All 7 answers →

  • gpt-oss-120b

    #6 · OpenAI

    OpenAI's open-weight model for reasoning and agentic tasks: 117B total and 5.1B active parameters, released with the smaller gpt-oss-20b in August 2025 under Apache 2.0.

    B5/7On the 7 tests
    • Is gpt-oss-120b under an open license?Yes
    • Can you download gpt-oss-120b's weights?Yes
    • Does the maker state gpt-oss-120b's context window?Yes
    All 7 answers →

  • NVIDIA Nemotron 3.5 Lightning

    #7 · NVIDIA

    NVIDIA's 30B-total, 3B-active hybrid Mamba-2, mixture-of-experts and attention model for agents and local inference, released 11 August 2026 in BF16, NVFP4 and GGUF under the OpenMDW-1.1 license.

    B5/7On the 7 tests
    • Is NVIDIA Nemotron 3.5 Lightning under an open license?Partly
    • Can you download NVIDIA Nemotron 3.5 Lightning's weights?Yes
    • Does the maker state NVIDIA Nemotron 3.5 Lightning's context window?Yes
    All 7 answers →

  • Kimi K3

    #8 · Moonshot AI

    Moonshot AI's open-weight native multimodal agentic model: 2.8T total and 104B active parameters with a 1-million-token context window, served on the Kimi Open Platform as kimi-k3.

    B−4½/7On the 7 tests
    • Is Kimi K3 under an open license?Partly
    • Can you download Kimi K3's weights?Yes
    • Does the maker state Kimi K3's context window?Yes
    All 7 answers →

  • Qwen3.8

    #9 · Alibaba (Qwen)

    The open release of Alibaba's Qwen3.8 generation: a 2.4T-total, 95B-active flagship (Qwen3.8-2.4T-A95B) and a 27B native vision-language model (Qwen3.8-27B), both on Hugging Face, with Qwen Cloud as the managed API.

    B−4½/7On the 7 tests
    • Is Qwen3.8 under an open license?Partly
    • Can you download Qwen3.8's weights?Yes
    • Does the maker state Qwen3.8's context window?Yes
    All 7 answers →

  • GLM-5.3

    #10 · Z.ai

    Z.ai's current flagship open release, a 753B-parameter model post-trained from the same base as GLM-5.2; GLM-5.3-Flash is the 320B-total, 18B-active multimodal model in the same series.

    C4/7On the 7 tests
    • Is GLM-5.3 under an open license?Partly
    • Can you download GLM-5.3's weights?Yes
    • Does the maker state GLM-5.3's context window?Not stated
    All 7 answers →

  • Muse Spark Open

    #11 · Meta

    Meta's open-weight model from Meta Superintelligence Labs, released 15 October 2026: 816B total and 42B active parameters with a 1M-token context window. It is Meta's newer open release than the Llama 4 models of April 2025.

    C−3½/7On the 7 tests
    • Is Muse Spark Open under an open license?Partly
    • Can you download Muse Spark Open's weights?Partly
    • Does the maker state Muse Spark Open's context window?Yes
    All 7 answers →

  • Olmo 3.1 32B

    #12 · Ai2 (Allen Institute for AI)

    Ai2's fully open 32B language model, released with its training data (Dolma 3, Dolci), code and checkpoints under Apache 2.0, in Instruct and Think variants alongside Olmo 3 7B.

    C−3½/7On the 7 tests
    • Is Olmo 3.1 32B under an open license?Yes
    • Can you download Olmo 3.1 32B's weights?Yes
    • Does the maker state Olmo 3.1 32B's context window?Not stated
    All 7 answers →
The tests

7 questions, the same for every open model.

  1. 1

    Is a open model under an open license?

    Yes means: The weights are under an OSI-approved license such as Apache-2.0 or MIT. Partly means the maker's own license with conditions, such as a user cap or an acceptable-use clause; no means research or non-commercial use only.

  2. 2

    Can you download a open model's weights?

    Yes means: The maker's page links the weights for download with nothing more than a click-through license. Partly means an access form, an approval step, or only some sizes released.

  3. 3

    Does the maker state a open model's context window?

    Yes means: A token count for the context window is on the maker's page or model card.

  4. 4

    Does the maker say what it takes to run a open model on your own machine?

    Yes means: The page states the parameter count and the memory or hardware it needs. Partly means the parameter count alone.

  5. 5

    Does the maker publish a price for its hosted a open model?

    Yes means: A per-token price on the maker's own pricing page. Partly means a hosted version exists but the price sits behind sign-in or a sales call.

  6. 6

    Does the maker train on your prompts when you use its hosted a open model?

    Yes means: Here no is the good answer: the maker's page says what you send its API is not used for training. Yes means it is unless you opt out; partly means it depends on the plan or the page does not give the default.

  7. 7

    Can you fine-tune a open model and keep the result?

    Yes means: The license allows fine-tuning and the fine-tuned weights are yours to keep and share. Partly means conditions follow the derivative, or a hosted fine-tune you cannot download.

Every grade is read from the maker's own documentation or pricing page on the day shown, with the page linked on the product's page. Nothing is graded from a review, a blog or a test we ran; where we test a product ourselves the page will say so. If a page has changed, tell us.

Questions people ask

The short answers.

Which open models are under Apache 2.0 or MIT?
Gemma 4, gpt-oss-120b and Olmo 3.1 32B are under Apache 2.0, and DeepSeek-V4.1-Flash, MiMo-V2.6-Pro and Phi-4-reasoning-vision-15B under MIT. Qwen3.8-27B, GLM-5.3-Flash and Mistral Small 4 are too, but their flagship siblings (Qwen3.8-2.4T-A95B, GLM-5.3 and Mistral Large 4) carry the maker's own license with a revenue threshold for running the model as a service, as do Muse Spark Open and Kimi K3. Nemotron 3.5 Lightning is under OpenMDW-1.1, a permissive license the OSI has not approved on the pages read.
Which makers publish a price for hosting their open model?
DeepSeek, Z.ai, Moonshot AI, Xiaomi, Mistral and Qwen Cloud list per-token prices on their own pages. Google serves Gemma 4 on the Gemini API without listing a Gemma price, Microsoft and NVIDIA host Phi and Nemotron without a per-token figure on the pages read, and Meta, OpenAI and Ai2 do not host Muse Spark Open, gpt-oss or Olmo themselves.
Which makers say they do not train on API prompts?
Xiaomi's platform policy says it does not use the text you send for training without your consent, and Microsoft says serverless deployments on Azure AI Foundry are not used to train models. Mistral and Google make it a plan or tier setting, Z.ai's API addendum says content is not stored without mentioning training, Moonshot's agreement says customer content may be used unless agreed otherwise in writing, and the rest say nothing on the pages read.

Placement here is never for sale and no link on this page pays us. How links work on this site. The consumer assistants are graded the same way on the Scorecard.