But Is It Open?The online edition
Type & theme

Make yourself comfortable

Page color
Text size
20 px
Reading width
Ask about this chapter ↗

Your settings and place stay in this browser.

Chapter 14 · Timothy O’Brien

A Field Guide to Open-Claiming Models

6 minute read22 source notesFree online edition

OpenRouter’s public rankings make a useful starting point for comparing models. They show token traffic through one routing service, rather than a global census of AI use. Self-hosted deployments and calls made directly to providers are outside that sample.1 Popularity can tell us which releases deserve attention. It cannot tell us their permissions.

The names on those pages mix downloadable weights, custom licenses, and proprietary services. A publisher may offer all three. This chapter compares concrete releases and then follows three differences that matter when choosing one: terms that split within a family, agreements attached to access, and the material needed to rebuild the work.

The comparison reflects public terms checked in November 2026. It is a guide to reading a release, not a standing approval for a future checkpoint. The companion dataset and website carry the profiles and their supporting evidence. A family overview can link several releases without assigning one score to incompatible licenses.2

A comparison of releases

Release or specified group Artifact license What needs attention
DeepSeek V4 Pro and V4.1 Flash MIT Weights and code are available; the full training corpus is not published.34
Z.ai GLM-5.3 Flash MIT Do not infer the same terms for the larger GLM-5.3.5
Z.ai GLM-5.3 Custom GLM-5.3 Qualifying large Model-as-a-Service operators need a vendor security review.6
Qwen3-8B Apache 2.0 This concrete checkpoint differs from later custom-licensed Max releases.7
Qwen3.8-Max, 2.4T-A95B Custom Qwen3.8-Max Display requirements and separate licensing for qualifying service businesses.8
Kimi K3.1-Pro Custom Kimi K3.1-Pro Service-business and display thresholds, with explicit exceptions for internal use and specified provider access.9
Mistral Small 4 119B (2603) Apache 2.0 This checkpoint uses different terms from Large 4 and modified-MIT releases.10
Mistral Large 4 Mistral Ecosystem Open-Weight Commercial Model-as-a-Service offerings require a separate license.11
Meta Llama 4 Scout and Maverick Llama 4 Community License User-count threshold, acceptable-use policy, branding, and an EU restriction on multimodal rights.12
OpenAI gpt-oss 120b and 20b Apache 2.0 Public weights; a different arrangement from OpenAI’s hosted frontier services.13
Google Gemma 4 Apache 2.0 Earlier Gemma generations used different terms; the training corpus and full recipe are not released.14
Ai2 Olmo 3 Apache 2.0 for released weights and project code Training, post-training, and evaluation resources make the release more reproducible; individual data resources need their own terms checked.15
Apertus 1.5 70B Apache 2.0 for released weights and project code The publisher’s download adds a separately accepted AUP; source datasets retain their own terms.16
EleutherAI Pythia Apache 2.0 for models and project code A research suite with intermediate checkpoints and reconstruction resources.17

Hosted systems such as GPT-6 Luna and Claude Opus 5.5 offer a different relationship: service access without downloadable weights or an independent runtime.18 They can still disclose useful technical information, so closed weights do not make every other dimension unscorable. They belong at the hosted end of the profile, with whatever evidence the other dimensions support.

One family name, different permissions

GLM-5.3 Flash uses MIT. GLM-5.3 adds a condition for a licensee or affiliate operating a Model-as-a-Service business when aggregate revenue exceeds $10 billion over any consecutive twelve months. Before commercial use, that operator must pass a security review whose scope and method Z.ai reasonably determines. The license excludes certain embedded end-user features and simple relaying from its service-business definition.65

A single “GLM” permission score would therefore answer the wrong question for at least one checkpoint. A conservative family score would understate the Flash grant; a permissive one would hide the larger model’s condition. Separate release profiles let the reader see both.

Qwen provides another example. The Qwen3-8B grant is Apache 2.0. Qwen3.8-Max’s custom text adds model-name display above specified user or revenue thresholds and a separate-license requirement for qualifying Model-as-a-Service or AI work-assistant businesses above $50 million in trailing twelve-month revenue.87 The publisher name and the download format are similar. The permissions are not interchangeable.

Mistral also publishes multiple regimes. Small 4’s 119B-2603 checkpoint uses Apache 2.0, while Large 4 adds a service-business condition and some modified-MIT releases carry revenue limits.101119 The practical rule is to record a model identifier and license revision, rather than approving “Mistral” as a category.

Kimi K3.1-Pro shows why a threshold summary still isn’t the whole clause. Its separate-agreement requirement covers qualifying Model-as-a-Service businesses above $20 million aggregate revenue over twelve months. Its model-name display threshold uses 100 million monthly active users or $20 million monthly revenue. The text then exempts internal-only use and access through specified official or certified providers from those sections.9 Someone running the weights locally for internal work can have a different question from a company selling model access.

The license and the download agreement

Apertus 1.5 70B is a particularly revealing comparison. Its weights are identified as Apache 2.0, but the official Hugging Face download requires the user to accept a separate AUP. That agreement requires an indemnity to ETH Zurich and EPFL and an express-assent chain on redistribution: recipients must accept the obligations and pass them on in turn.16

That deserves a Permission caveat. It does not mean Apache 2.0 has stopped being an OSI-approved license. The license on the artifact and the contractual promises accepted to obtain it are separate facts. Apertus’s trademark nonendorsement notice is separate again; Apache itself also reserves trademark rights.20

The training material needs the same care. Apertus publishes reconstruction code under Apache 2.0, but that code draws on external datasets. FineWeb, one documented input, is published under ODC-By and subject to CommonCrawl’s terms. An Apache license on a processing script does not relicense the corpus it processes.21

OpenAI’s gpt-oss illustrates a less restrictive combination. Its Apache grant sits beside a usage-policy file asking users to comply with applicable law; the license does not incorporate a separate prohibited-use list.13 Earlier Gemma terms did incorporate such restrictions, while Gemma 4 moved to Apache 2.0.14 The existence of an AUP or policy file is a prompt to read it, not a sufficient reason to classify every bundle alike.

What another group can rebuild

The remaining difference is what accompanies the weights. A runnable model can be useful without publishing the materials needed to reconstruct its training. DeepSeek’s current releases are permissively licensed and technically documented, including architecture and training information, while the corpus itself is not public.34 Those are substantial disclosures with a substantial limit.

Release permissions also do not settle training provenance. Allegations that DeepSeek distilled other labs’ systems concern how it obtained capabilities, a separate question from the rights it grants downstream.22 A field guide should preserve both questions rather than treating one as an answer to the other.

Olmo 3 goes further in publishing a development flow: data resources, intermediate checkpoints, training and post-training tools, and evaluation infrastructure. Those materials allow an independent group to investigate how the result was produced and attempt a comparable run.15 Reproduction still requires skill, compute, and attention to each resource’s terms.

Pythia was designed around research into training dynamics. Its original eight-size suite makes standard and deduplicated training runs available, with 154 checkpoints per model, training code, and resources for reconstructing the data order.17 It is not offered here as a substitute for the most capable current service. It shows what becomes possible when reproducibility is a design objective.

These comparisons are more useful than a publisher-wide verdict. Before choosing a release, identify the artifact you will actually use, the rights and agreements that follow it, and the evidence another person could inspect. The profile organizes those answers. It cannot supply a missing license or unpublished training corpus.

Notes

  1. OpenRouter, “AI Model Rankings,” https://openrouter.ai/rankings. Used here as an illustration of token traffic through the OpenRouter API, not as evidence of global market share or a fixed historical ranking. The live rankings change over time. ↩

  2. But Is It Open? companion dataset and methodology, edition snapshot dated November 15, 2026, https://github.com/butisitopen/data/tree/main/snapshots/2026-11-15; maintained release profiles, https://butisitopen.com/is-it-open/. Family overviews identify changing terms without assigning an aggregate tier to incompatible releases. ↩

  3. DeepSeek, DeepSeek-V4-Pro model page, https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro, and DeepSeek-V4.1-Flash model page, https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash (MIT for weights and accompanying code). The historical custom grant for original V3 must not be attributed to the MIT-licensed R1 release. ↩1 ↩2

  4. DeepSeek, V4 technical report, https://arxiv.org/html/2606.19348v1, and model-algorithm disclosure, https://cdn.deepseek.com/policies/en-US/model-algorithm-disclosure.html. These publish architecture and training information rather than the full corpus. ↩1 ↩2

  5. Z.ai, GLM-5.3-Flash MIT License, https://huggingface.co/zai-org/GLM-5.3-Flash/raw/main/LICENSE, checked November 15, 2026. ↩1 ↩2

  6. Z.ai, GLM-5.3 License, https://huggingface.co/zai-org/GLM-5.3/raw/main/LICENSE, checked November 15, 2026. The condition covers qualifying Model-as-a-Service operators above $10 billion aggregate revenue over consecutive twelve months; the definition excludes specified embedded end-user features and mere relaying. ↩1 ↩2

  7. Qwen team, Qwen3-8B Apache License 2.0, https://huggingface.co/Qwen/Qwen3-8B/raw/main/LICENSE, checked November 15, 2026. ↩1 ↩2

  8. Alibaba Qwen team, Qwen3.8-Max License, Hugging Face, https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/blob/main/LICENSE; Qwen3.5 license (Apache 2.0), Hugging Face, https://huggingface.co/Qwen/Qwen3.5-397B-A17B/blob/main/LICENSE; QwenLM, Qwen3.8 repository, GitHub, https://github.com/QwenLM/Qwen3.8. ↩1 ↩2

  9. Moonshot AI, Kimi K3.1-Pro License, https://huggingface.co/moonshotai/Kimi-K3.1-Pro/raw/main/LICENSE, checked November 15, 2026. Sections 2 and 3 specify different service-business and display thresholds; section 4 exempts defined internal use and access through official products or certified inference partners from those requirements. ↩1 ↩2

  10. Mistral AI, Mistral-Small-4-119B-2603 model card, https://huggingface.co/mistralai/Mistral-Small-4-119B-2603 (Apache 2.0); licensing overview, https://help.mistral.ai/en/articles/347393-under-which-license-are-mistral-s-open-models-available. Checked November 15, 2026. Apache status belongs to this checkpoint, not every release from the publisher. ↩1 ↩2

  11. Mistral AI, Mistral Large 4 open-weight announcement, https://mistral.ai/news/mistral-large-4-oss; Mistral AI, Mistral-Large-4-1T-A52B model page (Mistral Ecosystem Open-Weight License), Hugging Face, https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B. ↩1 ↩2

  12. Meta, Llama 4 Community License, https://dev.meta.ai/llama/llama4/license; Meta Llama organization (gated downloads), Hugging Face, https://huggingface.co/meta-llama. ↩

  13. OpenAI, “Introducing gpt-oss,” https://openai.com/index/introducing-gpt-oss; gpt-oss-120b LICENSE, https://huggingface.co/openai/gpt-oss-120b/raw/main/LICENSE, and USAGE_POLICY, https://huggingface.co/openai/gpt-oss-120b/raw/main/USAGE_POLICY, checked November 15, 2026. Apache 2.0 does not incorporate a separate prohibited-use list; the policy asks for compliance with applicable law. ↩1 ↩2

  14. Google, “Gemma 4 model card,” https://ai.google.dev/gemma/docs/core/model_card_4; Gemma Apache license, https://ai.google.dev/gemma/apache_2; earlier-generation terms, https://ai.google.dev/gemma/terms. The model card publishes architecture and evaluations; this does not amount to releasing the training corpus and complete training recipe. ↩1 ↩2

  15. Ai2, “Olmo 3,” https://allenai.org/blog/olmo3; Olmo-3-1125-32B model page, https://huggingface.co/allenai/Olmo-3-1125-32B. The release links training and post-training resources, checkpoints, code, and evaluations. Licenses for individual data resources should be read separately from the Apache grant on model and project code. ↩1 ↩2

  16. Swiss AI Initiative, Apertus-v1.5-70B model page and download assent, https://huggingface.co/swiss-ai/Apertus-v1.5-70B; “Apertus LLM Acceptable Use Policy,” v1.5, July 14, 2026, https://github.com/swiss-ai/apertus-legal/blob/main/apertus_1.5/USAGE_POLICY.pdf. The model page identifies Apache 2.0 and separately requires AUP acceptance. The AUP adds indemnification and downstream express assent. Checked November 15, 2026. ↩1 ↩2

  17. EleutherAI, Pythia project repository, https://github.com/EleutherAI/pythia, especially “Models,” “Reproducing Training,” and the release history; model card, https://huggingface.co/EleutherAI/pythia-12b. The original suite covered eight sizes in standard and deduplicated variants; smaller models and additional random seeds were added later. The repository describes the 154 checkpoints and data-order reconstruction resources. ↩1 ↩2

  18. OpenAI, “Introducing GPT-6 Sol and Luna,” https://openai.com/index/introducing-gpt-6-sol-and-luna; Anthropic, “Introducing Claude Opus 5.5,” September 22, 2026, https://www.anthropic.com/claude-opus-5-5. These describe service access without published weights and an independent runtime. ↩

  19. Mistral AI, model documentation, https://docs.mistral.ai/models; licensing overview, https://help.mistral.ai/en/articles/347393-under-which-license-are-mistral-s-open-models-available; Devstral 2 modified MIT example, https://huggingface.co/mistralai/Devstral-2-123B-Instruct-2512/raw/main/LICENSE (rights withheld when the specified monthly-revenue threshold is exceeded). ↩

  20. Swiss AI Initiative, legal notice, https://www.apertus-ai.org/pages/legal/ (trademark nonendorsement); Apache License 2.0, section 6, https://www.apache.org/licenses/LICENSE-2.0. The trademark notice is not part of the cited one-page Apertus v1.5 AUP. ↩

  21. Swiss AI Initiative, pretraining-data reconstruction repository, https://github.com/swiss-ai/pretrain-data (lists external input datasets, including FineWeb); Hugging Face, FineWeb dataset card, https://huggingface.co/datasets/HuggingFaceFW/fineweb, “Licensing Information” (ODC-By 1.0 and CommonCrawl terms). ↩

  22. Cybersecurity and Infrastructure Security Agency, “China-Based Artificial Intelligence Companies Conducting Industrial-Scale Distillation Campaigns Against U.S. AI Companies,” Cybersecurity Advisory AA26-251A, September 2026, https://www.cisa.gov/news-events/cybersecurity-advisories/aa26-251a. The advisory describes alleged campaigns; it is not a court finding. ↩

From But Is It Open? by Timothy O’Brien (Discursive, 2026). © 2026 Timothy O’Brien. All rights reserved.

Chapter 146 min leftNext →

The book, chapter by chapter.

Source note

Go to the notes & sources ↓

The vocabulary.

A short guide to the book’s recurring terms. Follow each chapter for the argument and its sources.

Open weights
A release makes a trained model’s learned parameters available. The accompanying terms decide what you can do with them. Access to weights alone does not provide the training data, code or recipe. Read the chapter →
Copyleft
A way of using license conditions to preserve recipients’ freedoms when covered software is passed on. The obligations depend on the license and on what you distribute or make available. Read the chapter →
Permissive licenses
Licenses such as MIT and Apache 2.0 grant broad reuse rights with comparatively few conditions. Notice obligations still apply, and a software license does not automatically cover a release’s data or other components. Read the chapter →
Reproducibility
The ability to reconstruct and examine how an artifact was produced using the resources that are published. Running an existing checkpoint is a different achievement from rebuilding its training. Read the chapter →
Governance
Who can participate, who makes decisions, and how authority over a project or release is shared. A right to use the code does not by itself create a voice in upstream decisions. Read the chapter →
Training data
The examples used to train a model. Publishing the data, documenting its provenance and describing how it was filtered are distinct kinds of disclosure; privacy and other rights may limit what can be shared. Read the chapter →

A passage worth sharing.

Select a passage in the chapter, or share the one you’re reading.