But is it open?

Tier 2 Restricted open-weight

The reference open music-generation stack, whose weights are non-commercial.

MIT for the AudioCraft code; CC-BY-NC 4.0 for the MusicGen model weights

The published MusicGen weights are non-commercial: CC-BY-NC 4.0 prohibits commercial use of the licensed material · Commercial deployment of MusicGen requires separate permission from Meta · Code and weights carry different licences in the same repository

The clearest case in the dataset of code openness and weight openness diverging inside a single release — you may run and modify it, but you may not sell what it makes.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Permissively licensed audio foundation weights from Moonshot.

MIT for the weights; Apache-2.0 and MIT for the code OSI-approved

Standard MIT and Apache-2.0 obligations only

Noteworthy because the surrounding Kimi text families carry custom terms: the audio branch shows the same lab choosing ordinary permissive licensing.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

An 82-million-parameter Apache-2.0 voice model that runs almost anywhere.

Apache-2.0 OSI-approved

Standard Apache-2.0 obligations only

Shows that permissive weights plus small size can matter more than transparency for real-world openness: this one actually ships into products.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Full-duplex speech dialogue with one of the most detailed training descriptions in the audio batch.

CC-BY 4.0 for model weights; MIT for the Python code and Apache-2.0 for the Rust backend — composite terms with no single OSI-approved software licence over the release

Weight licenses are content licences rather than OSI-approved software licences, and some Kyutai checkpoints carry non-commercial variants · The model card states the system is intended for research use and advises against professional or advice-giving deployment · Attribution is required for redistributed weights

Shows that the transparency axis can score well even when the weight licence is a content licence — openness here is a spectrum across A, P, D, T, and R rather than one switch.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Apache-2.0 speech recognition with a 0.6B model small enough for on-device use.

Apache-2.0 OSI-approved

Standard Apache-2.0 obligations only

Shows speech recognition quietly becoming a permissively licensed commodity, which changes the strategic position of commercial transcription APIs.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Open audio weights under the same community license that governs Stability's image releases.

Stability AI Community License

Organizations with more than $1M in annual revenue must contact Stability AI for enterprise licensing · Published releases carry a license tag of 'other' rather than a named OSI licence, and the repositories are gated by agreement · Stability AI has repeatedly revised the audio release terms

Extends the Stability AI pattern from images into audio, showing that one community-license design can cover a whole product line across modalities.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

The most widely deployed open speech recognition stack, under a permissive license.

MIT License (code and weights; model cards also carry Apache-2.0 metadata) OSI-approved

Standard MIT obligations only; no revenue threshold or use restriction

Demonstrates that permissive licensing plus a documented method can carry a model to universal deployment even without releasing the training data.

Full profile ↗ · snapshot 2026-11-07

Tier 5 Fully open / reproducible

AutoML with the search space exposed rather than hidden behind a service.

Apache 2.0 OSI-approved

Some bundled foundation-model backbones carry their own terms

The contrast with proprietary AutoML services is that the automation stays auditable and forkable.

Full profile ↗ · snapshot 2026-11-07

Tier 5 Fully open / reproducible

Automated model selection, fully open.

Apache 2.0 for open-source H2O-3; H2O-3 Secure (AutoDoc, SSO and RBAC, MOJO scoring into Hadoop and Spark, SLAs) is a closed commercial offering OSI-approved

Several enterprise features are closed in H2O-3 Secure

AutoML is the machine-learning equivalent of a published recipe, and the enterprise tier shows how openness gets carved back into features.

Full profile ↗ · snapshot 2026-11-07

Tier 5 Fully open / reproducible

MIT-licensed gradient boosting with no attachments.

MIT OSI-approved

Confirms the classical-ML openness baseline across independent implementations rather than a single project.

Full profile ↗ · snapshot 2026-11-07

Tier 5 Fully open / reproducible

The default Python ML library, permissively licensed and fully inspectable.

BSD 3-Clause OSI-approved

A decade of accumulated practice sits in a repository anyone can read and modify, the opposite of an API-only frontier model.

Full profile ↗ · snapshot 2026-11-07

Tier 5 Fully open / reproducible

Fully open by construction: the artifact is the method.

Apache 2.0 OSI-approved

The baseline the open-model debate is measured against; in classical ML, open is the default rather than a concession.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Same vendor, different license: Cohere 2026 releases are permissive Apache 2.0.

Apache 2.0 (North-Mini-Code-1.0, cohere-transcribe-03-2026, cohere-transcribe-arabic-07-2026) OSI-approved

Isolates the license variable - the openness of a release is decided by the terms attached, not by the brand.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Same lab, different licence: the code line ships under DeepSeek custom terms rather than the MIT used for V3 and R1.

DeepSeek License Agreement v1.0 (model licence), distinct from the MIT used by DeepSeek V3/R1

custom permissive-with-restrictions licence: redistribution allowed with the same use-based restrictions carried into derivatives, plus prohibited uses (illegal content, PII, defamation, discrimination, military use, fully automated decisions affecting legal rights)

Materially distinct from the DeepSeek text-family entry in the dataset; shows that openness must be read per model, not per vendor.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Code model whose training data is genuinely open, including the opt-out mechanism.

BigCode OpenRAIL-M

use-based restrictions with permitted-use and prohibited-use lists · commercial use allowed but conditioned on use restrictions

The strongest example of consent-aware open training data, and a direct answer to the claim that code corpora cannot be published lawfully.

Full profile ↗ · snapshot 2026-11-07

Tier 5 Fully open / reproducible

The closest thing in the batch to fully open and reproducible: code, data, weights and method.

MIT-style licence (custom text with an additional disclaimer); network weights published by the project for free use OSI-approved

Bundled third-party libraries under their own licences

The cleanest positive case, a frontier-adjacent game-playing system where nothing material is withheld.

Full profile ↗ · snapshot 2026-11-07

Tier 4 Deeply open

Open chess self-play at scale, with copyleft code and community-run training.

GNU GPL v3 for the engine, with an additional GPL section 7 permission for CUDA-linked builds; network weights distributed freely by the project OSI-approved

Copyleft obligations on redistributed engine code

A decade-long, still-running open training run, which is the strongest evidence that decision systems can stay open in practice.

Full profile ↗ · snapshot 2026-11-07

Tier 4 Deeply open

DeepMind's game-playing method survives without DeepMind's weights, in permissively licensed code.

Apache 2.0 for DeepMind mctx and pgx; MIT for alpha-zero-general and comparable projects OSI-approved

No official DeepMind MuZero weights exist; these are independent implementations

When the lab withholds weights, the method can still be reproduced, so openness can come from the field rather than the publisher.

Full profile ↗ · snapshot 2026-11-07

Tier 0 Proprietary service

The famous game-playing agent that was never released.

No released code or weights; paper and blog post only

Nothing published to license or reuse; it ran as an internal research artifact

The negative control, a landmark result whose openness stops at the paper, which is where open usually stops.

Full profile ↗ · snapshot 2026-11-07

Tier 0 Proprietary service

The open interface around a closed intelligence.

None published for the model — hosted service (early access)

No weights, no architecture publication, no technical paper, no license · Hosted only via console/API (also listed on OpenRouter as typesafe/jev-1.13 and jev-1.14) · Python and JavaScript SDKs and the workflow-eval harness ARE open source — the interface is open, the model is not

The sharpest new form of the book's question: a decision model whose SDKs, eval harness, and type guarantees are all inspectable while the model itself is a black-box service. 'Open source' is doing zero work here — the openness lives entirely at the boundary.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

The community's answer to a closed decision model: the idea, rebuilt, minus the intelligence.

Apache 2.0 OSI-approved

Random weights — explicitly not the production Jev model · Makes no claim of matching TypeSafe's published results

The paired control to Jev: when a category-defining model stays closed, the open ecosystem reconstructs the interface — and the reconstruction is honest that it is not equivalent. Interface compatibility is not model equivalence.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Permissively licensed embedding family that is the de facto default for retrieval.

MIT for bge-m3 and the v1.5 encoders; Apache 2.0 for bge-reranker-v2-m3 - verify per checkpoint OSI-approved

Embeddings carry the same openness questions as generation models, but the licence answer here is unusually clean.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

MIT-licensed embedding baseline still widely used for retrieval.

MIT (multilingual-e5 series, e5-mistral-7b-instruct, mmE5-mllama-11b-instruct) OSI-approved

A second permissive embedding family, allowing the book to treat embedding openness as a solved licence problem with unsolved data problems.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Apache-2.0 embedding family spanning Qwen2 and ModernBERT bases.

Apache 2.0 OSI-approved

Completes the permissive-licence pattern across three independent embedding vendors.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Technically open embedding weights under research-only licences, including a corrected licence mistake.

CC-BY-NC 4.0 for v3; Qwen Research License for v4 (research-only, non-commercial; the model card states v4 was first published under CC-BY-NC 4.0 by mistake and corrected to the Qwen Research License because it derives from Qwen2.5-VL-3B)

non-commercial use only · commercial use requires a separate licence request to Alibaba Cloud under the Qwen Research License · attribution and Notice-file obligations on redistribution

Shows a licence error being corrected in public, and the practical consequence of inheriting a base model licence.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Apache-2.0 embeddings with Matryoshka truncation support.

Apache 2.0 OSI-approved

Same permissive pattern, and a useful example of flexibility features (variable dimensions) shipped as part of the open artefact.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

The open-weight image family whose own release table shows three different licenses in one row.

Mixed by variant: FLUX.1 [schnell] Apache-2.0; FLUX.1 [dev] FLUX.1-dev Non-Commercial License; FLUX.1 [pro] proprietary API only

The official model table lists different licenses per variant, so the family label hides the terms of the variant you actually download · FLUX.1 [dev] weights are released under a non-commercial license · FLUX.1 [pro] has no downloadable weights at all

A single name can carry Apache-2.0, a non-commercial custom license, and no weights at all — so openness has to be checked per variant, not per family.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Two variants with one genuinely permissive license and one non-commercial license.

Apache-2.0 for FLUX.2 [klein] 4B; FLUX Non-Commercial License for FLUX.2 [klein] 9B

The 9B weights are gated behind acceptance of the FLUX Non-Commercial License and the BFL Acceptable Use Policy · The smaller 4B variant is genuinely Apache-2.0, so terms differ within the family

A direct test of whether a family name means anything: the same 'open weights' label covers Apache-2.0 weights and non-commercial weights in the same release.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Top-of-chart open image weights with a revenue threshold written into the license.

FLUX Kommunity License v1.0

Non-commercial use only, except that Qualifying Users may generate commercial output if they have less than US$5M in gross annualized revenue · Acceptable Use Policy and trademark conditions apply · The Qwen3-VL-4B-Instruct text encoder inside the package is separately Apache-2.0

The clearest current example of 'open weights' plus a hard commercial ceiling: the license names a dollar figure, so the openness question becomes a business-model question.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

MIT-licensed image weights with a published technical report behind them.

MIT License (code and model weights) OSI-approved

Apply the standard MIT notice and copyright notice to redistributed weights and code

A useful counterweight to the FLUX and Stability families: the same distributable-weight practice with a genuinely permissive license attached.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

A strong Apache-2.0 image model whose successor reclassified itself as research-only and whose newest version has no weights.

Apache-2.0 for Qwen-Image and Qwen-Image-Edit; Qwen Research License Agreement for Qwen-Image-2.1

The 2.1 release (September 2026) is governed by a Qwen Research License Agreement rather than Apache-2.0 · Qwen-Image-3.0 ships no weights at all, so the newest capability is not downloadable · Permissions must be checked at each release, not at the family level

Shows how a permissive open-weight launch can be quietly replaced by a research license and an API-only roadmap — the label survives, the terms change.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

The most-deployed open image family, whose license tightened generation by generation.

CreativeML Open RAIL-M / Open RAIL++-M for 1.5 and SDXL; Stability AI Community License for 3 and 3.5

Use-based restrictions remain in the RAIL licenses, so they are not OSI-approved open source · Under the Community License, organizations with more than $1M in annual revenue must negotiate an enterprise license · Terms changed across generations inside the same family

Stability AI's own community license is the reference case for how an incumbent converts an open-weight release into a licensed-gated one without ever closing the weights.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Apache-2.0 image weights that trade capability for speed without adding license conditions.

Apache-2.0 OSI-approved

Standard Apache-2.0 obligations only: retain notices and state changes

Demonstrates that an OSI-approved weight license can coexist with essentially no training transparency — permissiveness and reproducibility are separate axes.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Open robot foundation model whose first weight releases are non-commercial.

Code Apache 2.0; N1 and N1.5 weights under the NVIDIA one-way non-commercial licence per their model cards; newer N1.7 weights under the NVIDIA Open Model License

N1 and N1.5 weights are non-commercial · NVIDIA Open Model License requires an attribution notice and Cosmos branding for Cosmos-derived work

Robotics repeats the text-LLM pattern: open code and checkpoints, with permission lagging capability.

Full profile ↗ · snapshot 2026-11-07

Tier 4 Deeply open

An open ecosystem layer: formats, datasets, training code and evaluation in one permissive repository.

Apache 2.0, with parts derived from Diffusion Policy under MIT OSI-approved

Individual Hub datasets and policies carry their own licences; check per repository

Shows that openness in robotics can be delivered as shared infrastructure rather than as a single open model.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Open code and weights for a robot policy, with Gemma-licensed weights and closed pretraining data.

Repository code Apache 2.0; checkpoints are PaliGemma and Gemma derivatives and carry Gemma terms rather than plain Apache 2.0

Weight terms are set by the Gemma licence, not the repo licence · The release is presented by the publisher as an experiment

Open robotics foundation models are real; the binding constraints are data provenance and backbone licensing, not access.

Full profile ↗ · snapshot 2026-11-07

Tier 4 Deeply open

Downloadable code, parameters and database, with share-alike databases and no training code.

Code Apache 2.0; parameters CC BY 4.0; mirrored databases mixed (BFD and UniRef30 CC BY-SA 4.0, MGnify and PDB CC0); AlphaFold DB CC BY 4.0

Share-alike obligations on BFD and UniRef30 mirrors · Attribution required for parameters and AFDB entries

The reference case for open in science: nearly everything needed to use and audit the system is downloadable, yet the training pipeline stays closed.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Open code, restricted weights: the sharpest licence contrast in the batch.

Code Apache 2.0; parameters under AlphaFold 3 Model Parameters Terms of Use plus the Model Parameters Prohibited Use Policy

Non-commercial use only, including research on behalf of commercial organizations · Weights may not be published or shared outside the requesting organization · Output may not be used to train similar structure-prediction models · Revocable licence; access can be terminated on breach

Same lab, adjacent year, radically different permission profile, which is exactly what the word open has to survive.

Full profile ↗ · snapshot 2026-11-07

Tier 4 Deeply open

MIT code and MIT weights over public sequence data: the practical open-weight baseline for biology.

MIT for source code and released checkpoints; ESM Metagenomic Atlas data under CC BY 4.0 with Meta Open Source Terms OSI-approved

Meta Open Source Terms apply to Atlas data use · Verify terms per checkpoint

Shows the protein-modelling openness ceiling that AlphaFold 3 weight terms fall short of, and the cleanest MIT-plus-public-data example in the batch.

Full profile ↗ · snapshot 2026-11-07

Tier 0 Proprietary service

Frontier protein model, API only.

Biohub platform terms and acceptable use policy; no published licence for flagship weights

API and platform access only · Acceptable Use Policy at biohub.org

Openness here is graded rather than binary: an open small model and a closed flagship under one roof.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

A genuinely open small checkpoint sitting beside a closed flagship.

MIT for code and the open checkpoint (Biohub/esm LICENSE.md and ESM3 README); the 2024 release shipped a non-commercial Community License requiring acceptance on the Hub OSI-approved

Earlier snapshots were gated behind non-commercial terms; check the exact release you cite

Illustrates the common pattern where openness is granted for a reduced model while the capable model stays behind an API.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

The original open-weight model, and the original missing-dataset problem.

Modified MIT License (adds an output-ownership paragraph; not verbatim OSI-approved MIT text)

Modified MIT: OpenAI disclaims ownership of generated content · WebText training data never released

GPT-2 set the template every later release followed: weights yes, data no — which is exactly why the rubric scores Access and Transparency separately.

Full profile ↗ · snapshot 2026-11-07

Tier 4 Deeply open

A small model trained entirely in public.

Apache 2.0 OSI-approved

Proof that the Chinchilla-optimal small-model recipe can be executed end-to-end on open data — the reproducibility ceiling most frontier releases never approach.

Full profile ↗ · snapshot 2026-11-07

Tier 5 Fully open / reproducible

Public-institution fully open model with strong provenance goals.

Apache 2.0 for source code, weights, and data, accompanied by a mandatory Apertus Acceptable Use Policy and an indemnity to ETH Zurich and EPFL

Hugging Face downloads require accepting an Acceptable Use Policy that adds obligations absent from Apache 2.0 · Users agree to indemnify, defend, and hold harmless ETH Zurich and EPFL against third-party claims · APERTUS is a registered trademark and may not be used to imply endorsement

Apertus is the sovereign and public-institution comparison to corporate open-weight releases: the only major non-commercial model family publishing its data, its recipes, and its evaluation protocol.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Open weights wrapped in a purpose-built responsible-AI licence.

BigScience RAIL License v1.0

custom Responsible AI License with a prohibited-use list (discrimination, medical advice, law-enforcement and justice-administration uses barred) · permissive grant modified from Apache 2.0

The origin point of the RAIL licence stack that later open-weight releases copied, including the pattern of attaching use restrictions to otherwise permissive terms.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Efficiency research released under Apache 2.0 but gated on unreleased kernels and an opaque base.

Apache 2.0 (HF license:apache-2.0 on Bonsai-8B-gguf, Bonsai-8B-mlx-1bit, Bonsai-8B-unpacked and Ternary-Bonsai releases) OSI-approved

Shows that openness of the artefact and openness of the training story can fail independently in the same release.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Research-grade open weights that are explicitly non-commercial.

CC-BY-NC 4.0 with Cohere Labs Acceptable Use Addendum (c4ai-cc-by-nc-license on Command R; CC-BY-NC-4.0 tag on Command A and Aya)

non-commercial only · Cohere Labs Acceptable Use Policy incorporated by reference · gated repositories requiring accepting the license agreement · use-based prohibited-use list

The clearest corporate example of research-only openness: usable, deployable, and legally closed to commercial use.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Permissive and influential open-weight family with substantial technical documentation and unresolved provenance debates.

MIT for the model weights and the accompanying code (deepseek-recipe, DeepInfer) OSI-approved

Earlier V3/R1-era repositories mix MIT code with a custom model license; verify each derivative

DeepSeek combines strong practical openness, permissive terms, efficient training claims, and controversy over possible distillation - capability, license, provenance, and allegations must stay separate.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Open-weight ERNIE exists under Apache 2.0, but the newest release arrives under a bespoke licence.

Mixed - Apache 2.0 for ERNIE-4.5-21B-A3B-Thinking and the ERNIE-4.5-VL checkpoints; custom ERNIE 5.5 Flash Base Model License Agreement v1.0 (dated 5 November 2026) for ERNIE-5.5-Flash-Base

the 5.5 licence is a custom use-restricted agreement with a prohibited-use list including military use, high-stakes automated decisions and professional practice without authorisation · definition of licensed materials explicitly excludes unreleased training code and data

Shows a large vendor moving from proprietary API-only to a split licence portfolio in a single year.

Full profile ↗ · snapshot 2026-11-07

Tier 4 Deeply open

Historically important model with meaningful but release-variable transparency.

Apache 2.0 for Falcon 7B/40B; TII Falcon LLM License 2.0 for Falcon 3 and the Falcon-H1 family

TII Falcon LLM License 2.0 is Apache-2.0-derived but adds an Acceptable Use Policy, a mandatory attribution notice, and commercial obligations · Hosted use of Falcon-180B may require a separate license · License varies by version and the OSI has disputed the open-source claim

Falcon is an early sovereign-AI comparison point and a clean case for why 'open source' has to be checked at the exact checkpoint, since one family spans plain Apache 2.0 and a custom license with conditions.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

A major lab family that moved to a standard license without opening the training system.

Apache 2.0 for the Gemma 4 weights, a change from the custom Gemma terms used for earlier generations OSI-approved

Earlier Gemma releases (1 to 3) shipped under a custom Gemma license with a prohibited-use policy and should be checked separately · The license does not cover Google trademarks, Google's API services, or the underlying training data

Gemma 4 is the clearest case of a standard software license on weights coexisting with an unchanged training system, and of a corporate lab answering the license debate without answering the reproducibility one.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Open-weight frontier coding model, not a full open-source AI release under the strict rubric.

Custom GLM-5.3 License for GLM-5.3; MIT retained for GLM-5.3 Flash and earlier GLM-5.x releases

Model-as-a-Service operators with more than $10B aggregate 12-month revenue must pass a Z.AI security review whose scope and method Z.AI determines · Release-specific terms; the license must be read rather than inferred from open branding

Z.AI explicitly uses the open-weight label and discusses vulnerability-discovery training, while the release provides vendor-controlled security review and no OLMo-level reproducibility - the label and the evidence point in different directions.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Highly usable open-weight model with Apache 2.0 weights but incomplete training openness.

Apache 2.0 on the weights plus a separate gpt-oss usage policy OSI-approved

A separate usage policy applies alongside Apache 2.0, though it does not condition the software grant

The gap between a permissive weight release and fully open AI, stated in OpenAI's own deliberately precise wording - and, as of late 2026, still the only open-weight release OpenAI has made, alongside the gpt-oss-safeguard models.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Frontier-scale open-weight family with custom or modified terms and incomplete training transparency.

Modified MIT for the K2 family; custom Kimi K3 License for Kimi K3

K3 requires a separate agreement with Moonshot for Model-as-a-Service businesses above $20M aggregate revenue over any 12 months · K3 requires the model name displayed prominently above 100M MAU or $20M monthly revenue · K2-family Modified MIT adds attribution and display conditions and a synthetic-data clause

Kimi combines frontier-scale capability and practical self-hosting with a license and transparency profile that is not the same as fully open source, and its 2026 scale makes the MaaS condition bite far more companies than it would for a small model.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Kimi Delta Attention model released under plain MIT, unlike the modified MIT of Kimi K2.

MIT (HF license:mit tag on Kimi-Linear-48B-A3B-Base and Instruct) OSI-approved

A useful within-family licence contrast for the same vendor, showing that branding the whole vendor open-weight would have been wrong.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Permissively licensed industrial open weights with a published training-content summary.

MIT (HF license:mit tag on Ling-2.6-flash, Ling-2.6-1T and their fp8/int4 variants) OSI-approved

A useful reference point for how a large non-Western lab describes its training data without opening the corpus.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Influential open-weight family with custom conditions, now partly superseded inside its own publisher.

Llama 4 Community License plus Acceptable Use Policy for Scout and Maverick; Muse Model License for Muse Spark Open; Apache 2.0 for Muse Glimmer 30B

Products above 700M monthly active users must request a separate Meta license, granted at Meta's sole discretion · Mandatory 'Built with Llama' attribution and a 'Llama' naming prefix for derivatives · Multimodal rights are withheld from individuals and companies based in the European Union · The Muse Model License requires a separate agreement for Model-as-a-Service providers above $20M gross revenue in any 12 months · Llama 4 repositories are gated on Hugging Face behind license acceptance

Llama is the essential case study in the politics of naming: it made open weights mainstream, its custom terms show why downloadable does not mean unrestricted, and its April 2026 retirement followed by an Apache 2.0 return under the Muse name in August 2026 shows the terms can change with the business strategy.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Frontier-scale MoE family whose openness tightens release over release, from modified MIT to outright non-commercial.

Modified MIT (M2); non-commercial MIT-style (M2.7); MiniMax Community License (M3) - verify per release

M2 adds a branding clause only for products over 100M MAU or 30M USD ARR · M2.7 permits non-commercial use only and requires prior written authorization for any commercial use · M3 grants non-commercial use and requires Built with MiniMax M3 display plus a notice or authorization above 20M USD yearly revenue

Shows that open weights can be shipped under a source-available reading of the word open, with the same vendor marking M2 open-source in a README while its own M2.7 LICENSE restricts commercial use.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

The permissive half of a portfolio whose openness depends on the exact release.

Apache 2.0 on Large 3, Small 4, the Ministral 3 family, Devstral Small 2, Magistral Small, and the Voxtral and Shieldstral models OSI-approved

Other releases in the portfolio carry different terms; check the exact model and license

Large 3, Small 4, and the Ministral 3 family are the clean Apache 2.0 counterweight to the rest of Mistral's lineup, and they make the family-level 'open' claim impossible to read without checking the checkpoint.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

European frontier-scale open weights with a commercial carve-out.

Mistral Ecosystem Open-Weight License, a modified MIT

Businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially · Attribution notice and license text must be retained in copies and substantial portions

Large 4 marks Mistral's flagship moving off Apache 2.0, and its MaaS carve-out joins the 2026 pattern of vendor-set revenue and business-model conditions attached to nominally open weights.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Open weights whose license switches off for larger companies.

Modified MIT with a company-wide monthly revenue threshold

If the licensee's consolidated global monthly revenue exceeded $20M in the preceding month, no rights under the license may be exercised and a commercial license must be sought · The threshold applies to total company revenue, not only to the service built on the model

Medium 3.5 is the sharpest test of the 'downloadable means unrestricted' assumption: the weights are open and the license is a modified MIT, yet the grant evaporates for any company above a revenue line.

Full profile ↗ · snapshot 2026-11-07

Tier 4 Deeply open

Commercial model family with unusually deep data and recipe disclosure.

NVIDIA Nemotron Open Model License for the Nemotron 3 line; OpenMDW-1.1 (Linux Foundation) for Nemotron 3.5 releases

OpenMDW-1.1 has been submitted to the OSI but is not yet approved · Terms are published per asset, so each checkpoint must be checked

Nemotron shows that a commercial hardware company can choose to publish beyond weights, and its move to OpenMDW-1.1 shows how a new permissive license family is being adopted for AI model distributions.

Full profile ↗ · snapshot 2026-11-07

Tier 5 Fully open / reproducible

Open research system: the reference case for end-to-end openness.

Apache 2.0 for weights, code, and tooling; training and fine-tuning datasets published for download without license restrictions OSI-approved

OLMo is the clearest counterexample to the claim that open training is impossible, and the only major family that publishes the whole model flow rather than a finished checkpoint.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

An enterprise AI company shipping genuinely permissive small models beside a proprietary platform.

Apache 2.0 (palmyra-base and the palmyra-mini family; verify per release) OSI-approved

Verify each checkpoint; the enterprise Palmyra X-series and domain models (Med, Fin) are API/proprietary

The sponsored-open-source pattern in miniature: a commercial vendor gives away small permissive artifacts that funnel toward a closed product line — and unlike Gemma, the licenses are clean.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Practical small and mid-size models with unusually good artifact disclosure for a commercial lab.

MIT across the Phi-4 family OSI-approved

Model cards state that nothing in the card is a restriction or modification of the license, but the release is version-specific in practice

Phi shows that smaller models can be more practically portable than frontier ones, and that a commercial lab can publish fine-tuning code and benchmark logs without opening the corpus.

Full profile ↗ · snapshot 2026-11-07

Tier 5 Fully open / reproducible

The cleanest true reproducibility case: same data, same order, every checkpoint published.

Apache 2.0 OSI-approved

The benchmark case for what fully open means, and the reason the book can show that reproducibility is a design choice rather than an accident.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Broadly deployable open-weight family whose top tier now carries a custom license.

Apache 2.0 across Qwen3, Qwen3.5, and Qwen3.8-27B; custom Qwen3.8-Max License for the 2.4T-A95B Max tier; older Qwen2.5 checkpoints retain a custom Qwen license

Qwen3.8-Max License requires displaying the model name above 100M MAU or $20M monthly revenue · Qwen3.8-Max License requires a separate Qwen license for Model-as-a-Service or AI Work Assistant businesses above $50M TTM aggregate revenue · Qwen2.5 checkpoints carry an MAU threshold and derived-model attribution conditions

Qwen shows how a model can be highly portable and permissively licensed at the small end while the frontier tier moves to revenue-conditioned terms, which is the clearest 2026 case of a family split across two license regimes.

Full profile ↗ · snapshot 2026-11-07

Tier 4 Deeply open

Small model with an unusually complete public engineering blueprint.

Apache 2.0 OSI-approved

Demonstrates that meaningful reproducibility at small scale is achievable without publishing every pretraining byte.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Same family, three different licences - Apache 2.0, CC-BY-NC and a branded Apache derivative.

Mixed across releases - Apache 2.0 for SOLAR-10.7B-v1.0 base; CC-BY-NC 4.0 for SOLAR-10.7B-Instruct-v1.0; custom Upstage Solar Licence (Apache 2.0 derived with a branding condition) for Solar-Open and Solar-Open2

licence changes between checkpoints · the Upstage Solar Licence permits commercial use but requires derivative models to be named Solar and to display Built with Solar

Retained as an entry with the explicit caveat that Upstage openness must be checked at the exact checkpoint and licence level.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Permissive licence undercut by a deployment dependency on unreleased upstream patches.

Apache 2.0 OSI-approved

Shows that a standard licence does not guarantee open deployment when the runtime lives in private forks.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Open video weights whose commercial use is a registration away from a monthly cap.

The CogVideoX License (with Apache-2.0 on the 2b checkpoint only)

Commercial use requires registering and obtaining a basic commercial license · Commercial use is capped at 1 million visits per month before a further commercial license must be negotiated · Military, illegal, and national-security-related uses are prohibited, and the license notes it may be replaced by a more comprehensive version · Governed by PRC law with Beijing Haidian District jurisdiction

Registration-plus-cap is a distinct licence design: nothing is hidden, but the right to serve commercially is metered.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Open video weights whose license does not apply in three of the largest markets.

Tencent Hunyuan Community License

The license does not apply in the European Union, the United Kingdom, or South Korea, and is expressly limited to a Territory that excludes them · Additional commercial terms kick in for licensees with more than 100 million monthly active users · A separate Acceptable Use Policy is incorporated as Exhibit A

Territory clauses are the least discussed form of restriction: a licence can be free and open everywhere except where the publisher decides it should not be.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Open video weights with a revenue threshold and a prohibited-use annex.

LTXV Open Weights License 0.X; LTX-2.x Community License for later releases

Entities with at least $10,000,000 in annual revenue are Commercial Entities that must obtain a separate paid Commercial Use Agreement · Attachment A lists prohibited uses, and unauthorized commercial use by a Commercial Entity carries liquidated damages · The training Data is explicitly not licensed under the agreement · Transfer of fine-tunes may require a paid license under the LTX-2.x Community License

The annex structure is the point: the license grants broad freedoms and then subtracts a long list of uses, which is how 'open weights' releases quietly reintroduce acceptable-use policy.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

An Apache-2.0 video model whose practical limits are technical, not legal.

Apache-2.0 OSI-approved

Standard Apache-2.0 obligations only

Useful because the reasons to use it are hardware and resolution rather than licence terms — openness is not the binding constraint here.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

The video model whose license file says Apache-2.0 and stops there.

Apache-2.0 OSI-approved

Standard Apache-2.0 obligations only; no revenue threshold, no territory carve-out, no field-of-use restriction

The useful control case for video: identical openness problems to its competitors, but no license caveat attached, which isolates the license variable from the transparency variable.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

MIT code around weights nobody ever licensed.

MIT (code); model card carries no license tag for the weights

Model card: 'out of scope' for all deployed use cases; no explicit weight license grant · Training corpus (WebImageText, 400M pairs) never released

The historical precedent for the whole weights-without-license pattern that later open-weight releases formalized.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

Open except for the checkpoints you actually want.

Mixed: Small = Apache 2.0; Base/Large/Giant = CC-BY-NC 4.0

Non-commercial on all but the Small variant

A permissiveness gradient inside a single release: the smallest model is free, the capable ones are not.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Fully permissive self-supervised vision backbone.

Apache 2.0 (code and released weights) OSI-approved

X-ray DINO weights: FAIR Noncommercial Research License · Cell-DINO code: CC-BY-NC

The earlier CC-BY-NC framing aged out — DINOv2 shows a lab re-releasing its vision line under a true open-source license.

Full profile ↗ · snapshot 2026-11-07

Tier 1 Source-available or access-limited

Gated weights under a bespoke agreement — open in marketing, licensed in substance.

Custom DINOv3 License

Gated HF access with manual approval and contact-info sharing · Royalty-free but non-transferable grant; redistribution only under the same agreement · Mandatory 'Built with DINOv3' attribution · Trade-control and military-use exclusions; Meta may amend terms unilaterally · No commercial-use clause found in the license text

The clearest vision-side example of openness retreating between generations: DINOv2 Apache, DINOv3 gated custom terms.

Full profile ↗ · snapshot 2026-11-07

Tier 2 Restricted open-weight

The image JEPA pioneer, under a non-commercial license.

CC-BY-NC 4.0 (code and weights)

Non-commercial use only

Shows the original JEPA release was open-access, not open source — the license moved before the architecture became famous.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

The community's answer to CLIP's missing license.

MIT (code); released weights inherit upstream dataset terms OSI-approved

Weights trained on LAION and similar datasets carry non-commercial research terms

Shows open source re-implementations fixing what the original release left closed — at the cost of inheriting dataset-licensing problems.

Full profile ↗ · snapshot 2026-11-07

Tier 4 Deeply open

The rare vision release where data, code, and weights all shipped.

Apache 2.0 (code, checkpoints, SAM 2.1 included) OSI-approved

SA-1B dataset under a separate Research License (non-commercial)

With SA-1B published, SAM is the vision line that comes closest to end-to-end openness — held back from Tier 5 only by the dataset's separate research license.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Infrastructure-grade world model under a promising-but-unapproved open license.

NVIDIA Open Model License / OpenMDW-1.1

OpenMDW-1.1 submitted to OSI license review 2026-08-14, not approved

The test case for OpenMDW: if OSI approves it, a whole class of NVIDIA releases moves from custom-terms to genuinely open source.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Open method, closed model: MIT code with no weights to open.

MIT OSI-approved

The public repo is the author's reimplementation, explicitly unrelated to Google or DeepMind · No pretrained weights released by any party

Stretches the rubric honestly: Access measures artifacts that exist, and for decision models the artifact is often the method itself — a different shape of openness than a weight release.

Full profile ↗ · snapshot 2026-11-07

Tier 0 Proprietary service

The frontier world model, fully closed.

None published — closed research preview

No weights, no code, no license · Genie 4 (Oct 2026) rolling out to trusted testers, then AI Ultra / Project Genie, then Gemini API / AI Studio

The tier-0 control for the world-model category: the most capable system in the space ships zero artifacts.

Full profile ↗ · snapshot 2026-11-07

Tier 3 Practical open-weight

Permissively licensed video world model — the JEPA line's most open release.

MIT and Apache-2.0 across released checkpoints (GitHub code split MIT/Apache-2.0) OSI-approved

Sibling jepa-wms research repo remains CC-BY-NC 4.0 · V-JEPA 2.1 checkpoints distributed via download links, not yet the Meta HF org

Contradicts the assumption that Meta's JEPA family is uniformly non-commercial: the world-model generation moved to MIT/Apache while I-JEPA and the research repos stayed CC-BY-NC.

Full profile ↗ · snapshot 2026-11-07