45 open-weight models, 53 closed ones and the licences between them

45 open-weight models, 53 closed ones and the licences between them

Of the 166 AI models on our board, 109 carry a recorded specification, and 45 of those link to downloadable weights. Twenty-two are Apache 2.0, eight are MIT, and the other 15 sit under community terms, non-commercial licences, an attribution-only licence or no stated licence at all. Another 53 are closed: proprietary, restricted, or reachable only through someone else's API. The row that shows the gap best is Mistral Large 4, which launched on 6 October with an API price, no licence and no weights link, and whose weights are promised by the end of October 2026.

None of these models is ours. We do not build, host or resell any of them, and no listing depends on a vendor paying us. Every figure below is a count of rows on our own board as of 7 October 2026.

How we counted

The board holds 166 published product rows filed as AI models, all with an active lifecycle status. Fifty-seven of them have no model specification recorded yet (mostly older, well-known models such as GPT-4 Turbo, Llama 3.1 405B, Whisper large-v3 and Stable Diffusion 3.5 Large), so we cannot say anything about their licences from our data and they are excluded from every count below. That leaves 109 rows.

A row counts as having downloadable weights when it carries a weights link. All 45 such rows also name a licence except one, so the licence families below come straight from the licence text we recorded. A row counts as closed when its recorded licence says proprietary or restricted and it has no weights link. That gives 53 closed rows. One more row, FLUX 3 Image, records a commercial weights licence but no link, so we hold it apart rather than call it open or closed. Ten rows have neither a licence nor a weights link, and one row with weights has no licence, which makes 11 rows with no licence recorded in total. The groups add up: 45 + 53 + 1 + 10 = 109.

Prices are the figures stored in each row, in dollars per million tokens unless stated. We have not filled any gap by guessing.

The 45 with downloadable weights

They come from 40 different providers, from IBM and NVIDIA to a single individual developer. Nine are mixture-of-experts designs recorded with total and active parameter counts, so a 27 billion parameter model and a trillion parameter one sit in the same group.

Apache 2.0: 22 rows

This is the largest family. The 22 are Agnes 3.0 Flash, Granite 4.2 8B, Granite 4.2 30B, Bonsai 2 27B, Needle 3, Laya, Light-O1-Preview, Kev, NASA-IBM Lunar Foundation Model, CLM-v0.1-8B, GLiNER2.5-Decide, Limite 1B Violetto, Lev, Intern-Decision, Clef, Clef-flash, AstaBrief 8B, Strands Decider 2B, Kolibri, Index-Translate-35B-A3B (preview), pplx-decider-v1-27b, EmbeddingGemma 2.

Granite 4.2 is IBM's pair, and the NASA-IBM Lunar Foundation Model is an image model rather than a language one. Cloudflare, Ai2, Aleph Alpha, AWS (through its Strands Agents project), Perplexity and Bilibili's Index team all appear here. Several of those are infrastructure or research organisations, which fits a pattern of releasing weights where the model is not the main product, though our rows do not say why they chose Apache.

MIT: 8 rows

The eight are Atria Dawn Preview, CUA-S1-FORMS, MiMo-V2.6-Pro, MiMo-V2.6-Flash, Ming-Image-0.1-Design, Ming-Image-0.1-Design-Layer, Naive-N0.5-Flash, apex-flash-1. The MiMo pair from Xiaomi are the heaviest: V2.6 Pro is recorded at about a trillion parameters in total with 42 billion active. CUA-S1-FORMS is the smallest on the whole board, a checkpoint of 2.8 MB, and its row records that the training code and dataset were published as well, which is the most complete release we hold.

Community and custom licences with conditions: 8 rows

The eight are Kimi K3, Altar-1, FLUX 3 Action, Nemotron 3 Diarization, ThinkingCap-Qwen3.8-27B, Darwin-180B-RSI, IQuest-Q1, Tinfield 1. The licence names tell different stories:

• Kimi K3's licence is recorded as "open weights, not OSI", so the label itself says it is not an open-source licence.
• Altar-1 uses the GLM-5.3 licence, because it is built on that model.
• Darwin-180B-RSI and Tinfield 1 use the Qwen Community Licence 1.0, a family licence that follows the Qwen base.
• IQuest-Q1 is recorded as modified MIT: commercial products must display the name "IQuest-Q1".
• ThinkingCap-Qwen3.8-27B is under the PolyForm Small Business licence, which by its name is aimed at small firms.
• FLUX 3 Action sits under the FLUX Kommunity licence and Nemotron 3 Diarization under the OpenMDW licence agreement.

Non-commercial: 3 rows

The three are LimiX-2, Qwen-Image-2.1, Hemmingway-1. LimiX-2 publishes weights under a named non-commercial licence, with use restricted to non-commercial work. Qwen-Image-2.1 is under the Qwen Research Licence, recorded as non-commercial only, with commercial use needing a separate licence. Hemmingway-1 is CC BY-NC 4.0. We hold no row whose licence text says "research only" in those words, so that group is empty today; the Qwen Research Licence is the nearest, and we have counted it as non-commercial because that is how the row describes it.

Attribution only, unspecified and missing: 4 rows

Pianissimo is CC BY 4.0, which asks for credit and little else. GLM-5.3-Flash is recorded only as "open weights" and Qwen3.8-27B as "Qwen (open weights)", so for those two we know the weights are available but have not recorded the exact licence name. Tev1-4B-experimental links to its weights but states no licence, the one row in the open group where a buyer has nothing to read.

Weights promised, API available now

Four rows describe weights that are not yet in anyone's hands, or not yet linked from our page. We count none of them as open.

• Mistral Large 4 launched on 6 October 2026 as a public preview with a recorded API price. Its row has no licence and no weights link. Mistral has promised weights by the end of October 2026. We list the API now and will add the weights, with the licence, when they ship. Until then it sits among the 11 rows with no licence recorded.
• Qwen3.8-Max has a licence line reading "proprietary (weights announced)". Weights are announced; no link exists yet.
• Step 5 Preview is recorded as proprietary, with BF16 open weights announced for 15 October 2026 and not published yet. That date is eight days after our count.
• GLM-5.3 is the reverse case. Its licence line says it was proprietary at launch and the weights were released about two weeks later. Our row still has no weights link, so by the evidence of our own text the row is behind the release, and its smaller sibling GLM-5.3-Flash is already in the open group. We will correct it when the link is verified.

A promise is not a licence. Until the files exist and the licence is published, a buyer cannot plan on self-hosting, and the date can move. Step 5 Preview is the cleanest test, because a single day on the calendar will show whether the announcement was kept.

The 53 closed rows

Closed means no weights link and a licence recorded as proprietary or restricted. Ten of them belong to Google and Google DeepMind together, eight to OpenAI, six to Anthropic and six to Alibaba across its two provider names, with Microsoft AI and xAI at three each. In all, 23 provider names appear.

The 53 are Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking, GPT Image 2.5 Flare, GPT Image 2.5 Sunburst, Claude Mythos 5.1, WeatherNext 3, Muse Spark 1.3, Gemini 3.8 Flash Cyber, Gemini 3.7 Flash, Gemini 2.5 Pro, Gemini 3.8 Flash, Claude Fable 5.1, GPT-6 Astra, o3, Claude Sonnet 5, GPT-4o, Qwen3.8-Max, Qwen3.8 Flash, GLM-5.3, DeepSeek V4 Pro, Grok 4.6, Claude Opus 5, Koa, Pareto 26.9, Qwen3.8-Omni-Flash, Step 5 Preview, GLM-5.3-FlashX, Grok 4.7, Solar Mini 4, Qwen3.8-Omni-Flash-Realtime, Grok Voice Transcribe 2.0, decision-machine-1, HappyOyster 1.0, GPT-6 Sol, GPT-6 Luna, Claude Opus 5.5, Gemini 3.8 Flash TTS, Gemini 3.8 Flash-Lite TTS, Ember-1, Qwen-Audio-3.1-Realtime-Plus, Perceptron Mk1.5, Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 4 Argon, Mercury Voice, Recraft V4.1 Flash, Embed 5 Pro, Embed 5 Fast, Eleven v4, Eleven v4 Turbo, MAI-Transcribe-2-Streaming, MAI-Voice-2.1, MAI-Voice-2.1-Flash.

Three rows are worth singling out. Koa is recorded as proprietary with the added note that Salesforce holds the weights and does not release them, which is a settled position rather than a pending one. Gemini 3.8 Flash Cyber is restricted to a named access programme, so even the API is not generally available. GLM-5.3-FlashX is a hosted tier of a family whose other members have weights, which shows that one family can hold open and closed rows side by side.

The same is true of Black Forest Labs. FLUX 3 Image carries a commercial weights licence, so the weights exist on commercial terms, but there is no public link on our row and we count it in neither group. The lab's FLUX 3 Action is open under a community licence, so one lab appears in two different places in this article.

Rows with no licence recorded

Eleven rows. Ten have no weights link either: Mercury 2.5, Kimi K2.8 Preview, SWE-2, Jev, Hy4 Preview, DeepSeek V4.1 Flash, Fugu Ultra v2.0, Fugu Max, Qwen3.8-Max-0902, Mistral Large 4. The eleventh is Tev1-4B-experimental, mentioned above. An empty licence field means we have not recorded a licence, not that none exists. Mistral Large 4 is the one with a dated promise. For the others, such as the DeepSeek, Kimi, Sakana and Tencent previews, we have not yet found a licence statement we can cite, and they stay in this group until we do.

Month by month: Apache arrives in October

Release dates run from May 2024 to 6 October 2026, but the board is bunched at the recent end. This table shows how many specified rows were released in each month, and how many of them have downloadable weights.

Release month	Specified rows	With weights	Share
Before July 2026	4	0	0%
July 2026	2	1	50%
August 2026	11	4	36%
September 2026	79	32	41%
October 2026 (to the 6th)	13	8	62%

The four early rows are GPT-4o, Gemini 2.5 Pro, o3 and Claude Sonnet 5, all closed. The only July open row is Kimi K3. The shift in the licence mix is sharper than the shift in the share. The four open August rows are two Granite models under Apache 2.0 and two with an unspecified open-weights licence. September's 32 split into 12 Apache 2.0, 8 MIT, 7 community or custom, 3 non-commercial, 1 CC BY and 1 with no licence. All eight October rows are Apache 2.0.

Be careful with what this shows. The board is not a census of the market. It is what we have found and entered, and September is overrepresented because that is when we did most of the entering. The October figure covers six days and 13 rows, so one or two additions would move the share by several points. What we can say is that, among the rows we hold, the permissive licence has gone from one of several families to the only one in the latest week.

Prices: five open rows, 36 closed

Prices are the weakest part of the board, and the gap is itself a finding. Of the 45 open rows, five carry a price: Agnes 3.0 Flash, MiMo-V2.6-Flash, MiMo-V2.6-Pro, Naive-N0.5-Flash and Hemmingway-1. Of the 53 closed rows, 36 do. That is 11% of the open group and 68% of the closed.

The likely reason is plain. A closed model is sold by its maker, so there is a price page. An open model is often just a file, and the cost of running it depends on whoever hosts it. The five open prices we hold are prices recorded for API access, not for the weights, which cost nothing to download under their licences but cost hardware to run.

What the recorded figures say, with the caveat that the groups are not alike:

• Open, five rows: input prices run from $0.05 to $0.435 per million tokens, with a median of $0.14, and output prices from $0.15 to $0.90, with a median of $0.40.
• Closed, 35 rows with a per-million-token input price: from $0.04 to $10, with a median of $1.25. Among the 33 that also record an output price, the median is $6 and the highest is $50.
• Seventeen of the 35 closed input prices are $2 or more. None of the five open ones is.
• Nine closed rows have an input price under $0.50, so cheap closed models exist. decision-machine-1 is recorded at $0.04, Solar Mini 4 and GPT-6 Luna at $0.10.

The closed 35 include flagship models at $10 for input, such as Claude Mythos 5.1, Claude Fable 5.1 and GPT-6 Astra, while the open five are small or mid-sized models. This is not a like-for-like comparison, and we would not conclude from it that open weights are cheaper. The narrower reading is that where an open model has a price on our board, it sits at the low end of the range, and for most open rows no price exists to compare. Some prices carry conditions such as launch discounts or preview pricing, and the row notes keep those.

We covered how token prices are built in what makes a token expensive, and the evidence problem for small models in small models, winning argument, losing evidence. Neither repeats the counts here.

What each licence family lets a business do

This section is general guidance drawn from what the licence names imply. It is not legal advice and it is not a reading of any specific text. Read the actual licence for the model you plan to use, and ask a lawyer if the use matters to your revenue.

Apache 2.0. A permissive licence. In general it allows commercial use, modification and redistribution, and it includes an express patent grant. You normally have to keep the licence and notices, and state what you changed. It has no revenue or user-count threshold. This is the family a procurement team can usually approve fastest, and it covers 22 rows.

MIT. Shorter and more permissive still. Commercial use, modification and redistribution are generally allowed provided the copyright notice and licence text travel with the software. There is no patent clause, which some legal teams notice. "Modified MIT" is a different animal: IQuest-Q1 adds a display condition for commercial products, so it must be read, not assumed.

Community and custom licences with conditions. The word "community" in a licence name tends to signal that the owner wants wide use but keeps the right to draw a line. Typical lines in licences of this kind are a threshold on monthly users or revenue above which a separate agreement is needed, a ban on using the model's output to train a competing model, naming or attribution rules, and acceptable-use policies. Redistribution of modified weights may be allowed with the same terms attached. We have not recorded the thresholds for the eight rows, and a buyer should check each one. A licence labelled "not OSI", as Kimi K3's is, signals that it does not claim to meet the Open Source Definition.

PolyForm Small Business. The name implies a limit by the size of the user. In general, licences of this type permit use by small organisations and require larger ones to buy a different licence. Check the definition of "small" before relying on it, since it usually combines headcount and revenue. ThinkingCap-Qwen3.8-27B is the one row here.

Non-commercial licences (CC BY-NC and similar). These allow study, testing and personal or academic use, and they forbid commercial use. A company evaluating one internally may be fine, but putting its output into a paid product generally is not. For Qwen-Image-2.1 the row says commercial use needs a separate licence, which is a common pattern: the free version is a trial of the paid one.

CC BY 4.0. Allows commercial use and adaptation with credit to the author. It is a content licence more than a software licence, and Pianissimo is the only row that uses it.

Research-only. Normally means no deployment in production. We hold no row labelled in that way today.

Commercial weights licence. FLUX 3 Image's label suggests a paid licence under which you receive files. It is open in the sense that you may run it yourself, and closed in that you pay for it.

Three practical points follow. First, check the licence of the base model: Darwin-180B-RSI, Tinfield 1, ThinkingCap-Qwen3.8-27B and Altar-1 all inherit conditions from a parent model. Second, licence and weights are separate things: a model can be Apache licensed with weights hosted behind a sign-up wall, and the reverse. Third, a licence applies to a version. A later release of the same family can move from permissive to restricted, or from proprietary to open, which is what the GLM-5.3 row describes, so record the licence version alongside the model version in your own approval files.

Open weights are not open source

Almost every licence in the 45 is about the weights, the trained numbers. That is not the same as the open-source definition, which concerns the freedom to study, modify and redistribute the source from which something is built. For a model, the source would include the training data, the training code and the recipe. Most of these releases give you the weights and sometimes inference code, and nothing else.

A buyer should care for practical reasons. Without the training data you cannot reproduce the model, audit it for material you are not licensed to use, or judge what it was exposed to. Without the training code you cannot retrain it. You can fine-tune and run it, which is real freedom, but you inherit its behaviour as it is. CUA-S1-FORMS records that its training code and dataset were published, and that is why we picked it out earlier. It is the exception on our board.

There is also a difference between a licence the Open Source Initiative recognises and one it does not. Apache 2.0 and MIT are on that list. A custom community licence with thresholds is not, and Kimi K3's label says so itself. For a firm with a policy that allows only open-source software, the label matters, and "open weights" in a vendor's announcement does not meet it by default.

Finally, the two terms hide a third category: weights you may download but only under terms that exclude your use. Three non-commercial rows are in that position for most companies.

Limits of this count

The 57 rows without a specification are a real gap, and several of them (Llama 4 Maverick, Qwen2.5-72B, DeepSeek-R1, Gemma 2 27B, Phi-4) are well-known models whose licences are widely discussed. Adding them would likely raise the open count and change the month table, which is why we show 109 rather than 166. The licence text we store is a short label, not the licence, and we have not read each full agreement. Where a licence line is blank we say so rather than infer it.

We have written before about how little independent scoring exists and about what happens when the model you build on stops answering. The licence is the other half of that second risk: a model with open weights under Apache 2.0 can be kept running after its maker withdraws the API, while a closed one cannot. The board shows that choice is available for 45 of 109 specified rows today, and that four rows, Mistral Large 4, Qwen3.8-Max, Step 5 Preview and GLM-5.3, are the ones to watch.