Open weights are not open source: the three licence states behind 37 models

Three things to remember

  1. On this site, openSource means only that weights are downloadable; the licence field determines commercial use, modification and redistribution.
  2. There are 23 Apache-2.0/MIT models, 10 community licences and 4 research licences; the community group alone involves at least seven sets of terms that cannot be inferred from one another.
  3. Seven checks before procurement: version, commercial scope, redistribution, derivatives, attribution and registration, patents and trademarks, outputs and data.

Thirty-seven models in the seventh edition have downloadable weights. It is tempting to read that as “37 open-source models”. It is not.

“Weights downloadable” is only the first layer. Whether you can use them commercially, modify them, redistribute them, rename a derivative, or rely on a patent grant depends on the licence, not on the download button. This essay splits the 37 into three states and gives a pre-procurement checklist you can actually inspect.

Three words, three sets of rights

Open weights: the parameters are publicly obtainable. This describes distribution, not usage rights.

Open source: under the OSI definition, the licence must allow free use, modification and redistribution without field-of-use restrictions. Apache-2.0 and MIT are common examples.

Open API: only a callable endpoint, no weights. It is outside this essay, and it is the state of the 19 closed models on the board.

The openSource field on this site means only the first layer: weights can be downloaded. The licence field is the second layer. One cannot substitute for the other, and the word “open source” should not blur them.

The 37 in three states

State Models Licence Examples Default answer on commercial use and redistribution
Permissive open source 23 Apache-2.0 (18), MIT (5) Qwen3.8-VL-Flash, Qwen3.8-27B, DeepSeek-V4.1-Flash, GLM-5.3-Flash, Intern-S2, Wan2.2 Usually yes, with copyright and licence notices preserved
Community licence 10 Vendor-specific Kimi K3, MiniMax M3, MiMo-V2.6-Pro, Step 5 Preview, GLM-Image, HunyuanImage 3.0 Read clause by clause; commercial conditions or registration often apply
Research licence 4 Research use only LongCat-Next, Emu3.5, BAGEL-2, Zidong Taichu 2.0 No commercial use by default unless separately licensed

Strictly speaking, only the first 23 are close to “open source” in the usual sense. The other 14 are open-weight, not open-source. This site used to call all 37 “open-source models”; from the seventh edition we say “open-weight” and keep a licence summary on every model page.

Permissive is not obligation-free

Apache-2.0 and MIT are often treated as “use it however you like”. That is a misunderstanding. They do allow commercial use, modification, redistribution and self-hosting, but with conditions:

Obligation Apache-2.0 MIT
Keep copyright and licence notices Required Required
Include a copy of the licence when distributing Required Required
State modified files Required Not required
Patent grant Explicit Not explicit
Trademark grant No No
Field-of-use restriction None None

For most enterprises the real risk is not “can we use it” but “did we mark modified files as Apache-2.0 requires” and “did we ship the NOTICE file”. Those are the two items audits most often miss.

One label needs its own line: MiniCPM-V 4.5 is Apache-2.0 but the vendor asks for commercial registration. Our licence column says “Apache-2.0 (registration required for commercial use)” so readers do not stop at the first half.

Community licences: the widest spread

The 10 community-licence models involve at least seven different sets of terms: Kimi K3, Xiaomi MiMo, MiniMax, StepFun, Tencent Hunyuan, Zhipu GLM-Image and Kuaishou Keye-VL. Their only common feature is “weights downloadable and some commercial use allowed”. Specific differences can include:

  • separate permission above a monthly-user or revenue threshold;
  • attribution or registration when offering the model as a service;
  • a ban on training similar models from its outputs;
  • a requirement to keep the same licence when redistributing;
  • additional permission for certain jurisdictions or industries.

Those terms do not always appear together, and one model’s terms cannot be used to infer another’s. There is no industry-standard community licence; each one has to be read. Our summaries are for orientation, not a substitute for the original.

Research licences: commercial use is off by default

The four research-licence models — LongCat-Next, Emu3.5, BAGEL-2 and Zidong Taichu 2.0 — publish weights but limit use to research, evaluation or non-commercial experiments. Their scores are not low: Emu3.5 leads generation at 98. That does not change the licence boundary.

If your team wants to use Emu3.5 for commercial content generation, the correct path is to obtain a commercial licence from the vendor, not to ship because “the weights are on Hugging Face”. A download URL is not a permission letter.

One easy-to-confuse example

Kimi K3 is open-weight under a community licence, grade A evidence, 95.3 composite, ranked #2. It is strong, downloadable and commercially usable within its terms — but it is not OSI open source. Blurring those three leads to two opposite mistakes in procurement:

  • Legal assumes “open source = Apache-2.0” and rules Kimi K3 unusable;
  • Engineering assumes “downloadable = open source” and treats Kimi K3 as freely redistributable.

Both are wrong. The accurate description is: open weights, community licence, commercial use subject to the terms.

Seven checks before procurement

  1. Modality and version: confirm the licence matches the exact weight version you downloaded, not another model in the family.
  2. Commercial scope: whether you may offer it as a service, and whether industry, revenue or user thresholds apply.
  3. Redistribution: whether you may pass weights or fine-tuned weights to customers, subsidiaries or third parties.
  4. Derivatives: whether distillation, fine-tuning and merging are allowed, and whether a derivative may change licence.
  5. Attribution and registration: whether notices, source attribution, commercial registration or filing are required.
  6. Patents and trademarks: whether there is an explicit patent grant, and whether the vendor’s marks may be used in marketing.
  7. Outputs and data: who owns outputs, and whether using them to train a competitor is prohibited.

Put the seven into one table and have legal and engineering sign it. That is more productive than arguing in a review meeting about whether something “counts as open source”.

How this site labels it

  • openSource: whether weights are downloadable, true or false.
  • Licence column: Apache-2.0, MIT, community, research, plus conditions such as commercial registration.
  • Evidence grade: how much independent support the score has; unrelated to licence.
  • Composite: the eight-dimension weighted result; it contains no licence preference.

We do not add points for a permissive licence or remove a model for a strict one. The board answers “how capable is it”; the licence answers “may you use it this way”. A complete selection exercise answers both.

This essay is information, not legal advice. Before commercial deployment, have legal check the original licence and its current version.

FAQ about this essay

Does Apache-2.0 allow unconditional commercial use?

Commercial use is allowed, but not unconditionally: you must keep copyright and licence notices, include a licence copy when distributing, mark modified files and carry the NOTICE file where applicable.

Can research-licence models be used commercially?

No by default. Downloadable weights are not a commercial licence; contact the vendor for separate permission before deploying.

CN Multimodal Tier List · Editorial desk — An independent review project, not affiliated with or sponsored by any vendor. Weights, method and evidence grades are public. Report factual errors via the about page; corrections are logged.

Cite as: CN Multimodal Tier List. (2026). Open weights are not open source: the three licence states behind 37 models. 2026cnmllmtierlist.com