Open Weights in 2026: Chinese Labs Ship the Frontier, the Licenses Grew Clauses, and Brussels Started Enforcing

Once a quarter I decide which open-weight models my platform will offer as a one-click deployment, and every time the decision has got harder in a way that has nothing to do with the models. The models are excellent. The problem is that "open weights" now covers at least six legally distinct things, that the most capable ones come with revenue thresholds that are easy to miss, and that since 2 August a European regulator can fine you for shipping one without the paperwork. So this time I read every licence file rather than the model card, and I read the regulation rather than the summaries of it. This is what I found, as of 8 September 2026.

The short version: the open frontier is Chinese, MIT-licensed at the base, and roughly eight points and three to six months behind the closed frontier. Meta left the game in April and came back in August under Apache 2.0. Four of the top Chinese labs have added revenue-gated clauses that turn "MIT" into "MIT until you are big". And the EU's general-purpose AI rules were not delayed by the omnibus that delayed everything else, so they are the one part of the AI Act with teeth right now.

Who releases what

Here is the state of the leaderboard-relevant open models. Parameter counts are total / active for the mixture-of-experts models. I have verified licences from the licence files on Hugging Face, not from press coverage, because as I will show the two disagree.

Some things jump out. Five labs, all Chinese, have shipped models over 700 billion parameters this year, and Hugging Face's own Summer 2026 report notes that in five of the seven months of 2026 the largest American open release was under 130B. The exceptions are Arcee, a small company that spent a reported $20 million on a 33-day run on 2,048 B300s to produce a 399B model, and NVIDIA, whose Nemotron Ultra is promised at around 500B. OpenAI has not followed gpt-oss with anything in thirteen months. And Qwen is the download leader by a margin that is not close: 2.05 billion downloads against Google's 418 million and Meta's 227 million, with 151,448 derivative models on the hub.

Then there is Meta. Llama 4 in April 2025 was received badly, partly because the LMArena entry that ranked second was an "experimental" chat-tuned variant that was not the downloadable checkpoint, and the 2-trillion-parameter Behemoth never shipped. The meta-llama organisation on Hugging Face has not created a new model repository since 28 April 2025; I checked the API this morning. After the $14 billion Scale AI deal and the formation of Meta Superintelligence Labs, the first model out of the new group, Muse Spark in April 2026, was closed and API-only, the first Meta frontier model without weights since before Llama 2. Then on 10 August Meta released Muse Glimmer, a 30B model distilled from Spark, under Apache 2.0 in a new meta-models organisation, and said the weights of Spark 1.2 would follow. Zuckerberg's line was that "our goal should be for American open source models to be the best globally." That is a different goal from the 2024 one, which was for Meta's models to be the best. It is also, as far as I can tell, the end of the Llama brand and of the Llama Community License for new models.

The licence ladder

"Open weights" now means one of these, in decreasing order of freedom:

1. Plain Apache 2.0 or MIT. DeepSeek, Gemma 4, gpt-oss, Mistral Large 3 and Small 4, Xiaomi MiMo, Granite, Phi-4, Qwen's smaller models, Arcee. You can do anything, including serve it as a product and train on its outputs.
2. OpenMDW. The Linux Foundation's Open Model, Data and Weights licence, version 1.1 from May, which NVIDIA has adopted across Nemotron, Cosmos, GR00T and Isaac. Permissive, and unusually it covers the data as well as the weights, which is what the OSI's definition requires.
3. MIT plus attribution above a scale threshold. Moonshot introduced this with Kimi K2.5 in January: if your product exceeds 100 million monthly active users or $20 million in monthly revenue, you must "prominently display" the model name in the interface.
4. MIT plus a separate deal above a revenue threshold. This is the 2026 innovation, and three of the four largest Chinese releases carry it. From the Kimi K3 licence: model-as-a-service operators with more than $20 million in revenue over any twelve months "must enter into a separate agreement with Moonshot AI". From the Qwen3.8-Max licence on the 2.4T model: a model-as-a-service or "AI work assistant" business over $50 million in twelve months "shall obtain a separate license from Qwen", though purely internal use is exempt. And from the GLM-5.3 licence: a model-as-a-service operator over $10 billion must "pass Z.AI's security review before using the Software or its derivative works for any commercial purpose", with the scope of that review "reasonably determined by Z.AI".
5. Revenue-gated outright. Mistral Medium 3.5's Modified MIT simply does not license companies with more than $20 million in monthly consolidated revenue. MiniMax M2.7 in April went further and shipped non-commercial-by-default, with commercial use needing written permission; the licence was reportedly changed hours before release, and I could not fetch the original text, so treat that one as reported.
6. Community licences. Llama 4's: over 700 million MAU on the release date needs Meta's discretionary permission, "Built with Llama" must be displayed, and any model fine-tuned from it must have a name starting with "Llama".

The press coverage of tier 4 is unreliable in both directions. The Decoder reported the 2.4T Qwen under Apache 2.0; it is not, only the 27B is. Several outlets reported a "30% revenue share" in the Kimi K3 licence; there is no percentage in the text. The Hugging Face API is more reliable than either, because the license: other tag is what the labs themselves set when the licence is custom. This is what I ran:

# which Qwen3.8 checkpoints are actually Apache?
curl -s 'https://huggingface.co/api/models?author=Qwen&search=Qwen3.8&limit=20' \
  | jq -r '.[] | "\(.id)\t\(.tags[] | select(startswith("license:")))"'
# Qwen/Qwen3.8-27B            license:apache-2.0
# Qwen/Qwen3.8-2.4T-A95B      license:other
# Qwen/Qwen3.8-Flash-Next     license:other

# and read the custom text itself rather than the model card
curl -sL https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/LICENSE | grep -n -i 'revenue\|separate license'

To be clear about the practical effect: for a company under $20 million a year, none of the tier 3 and 4 clauses bind, and the models are as free as MIT. For a company at OpenRouter's or Together's scale, the four biggest open models each require a negotiation with a Chinese lab, and the GLM clause hands Z.AI a discretionary review over any commercial use. That is a very different licensing position from 2024, when the question was only whether you were bigger than 700 million users. The labs are not being unreasonable, exactly; they are pricing the thing everyone said was free. But "MIT" in a headline is no longer sufficient information.

None of these are open source by the OSI's definition, which requires enough data information to rebuild the model and no field-of-use restrictions. The models the OSI actually lists as conformant are Pythia, OLMo, LLM360 and T5. AI2's Olmo 3, released in November with every checkpoint, the Dolma 3 data and the training code, is the only frontier-adjacent release that clears that bar, at 32B. Nemotron 3 comes close with 3 trillion tokens of released training data. Everything else in the table is open-weight, and I would like the industry to say so.

How far behind is the open frontier

Depends on the index, and cross-index comparisons are meaningless, so here is one index at one time. Artificial Analysis Intelligence Index v4.3, this week: Claude Fable 5.1 and GPT-6 Astra tie at 53; the best open-weight models are GLM-5.3 at 45, Kimi K3 at 44, GLM-5.3-Flash at 42 and Qwen3.8-2.4T at 40. Eight points. On LMArena the top ten are all proprietary and the best open model sits around 30 Elo below the top, though the aggregator snapshots I found disagree on whether that is GLM-5.3-Max or Kimi K3, so I will not quote a rank. On SWE-bench Verified the self-reported open numbers are MiniMax M2.5 at 80.2%, Qwen3-Coder-Next at 70.6%, DeepSeek-V3.2 at 70%, against mid-70s to 80s for the closed frontier, and DeepSeek's own V4-Pro GA checkpoint jumped from 12.8 to 62.7 on DeepSWE between preview and release, which tells you how much of the gap is post-training rather than scale.

OpenRouter's June analysis put the lag at three to six months and added that "the frontier labs do not appear to be accelerating away." Their own token data through November 2025 had open models at about a third of usage, with Chinese open models at 13% for the year and 30% in some weeks. I have seen claims that open weights crossed 46% or a majority of OpenRouter tokens by mid-2026; I could not get the rankings page to return data and the secondary numbers contradict each other, so I am not repeating them. A third of one large router's traffic is the verified figure.

Where the gap is not closing is at the top of the price list. DeepSeek V4-Pro at peak is $1.32 in and $3.96 out per million tokens; GLM-5.2 is $1.40 and $4.40; Mistral Large 3 is $0.50 and $1.50. The closed frontier is $10 to $50 out, as I tabulated in my inference post. The eight-point gap costs ten times as much, which is the entire commercial argument for the open column.

The distillation fight and the chips

Two things happened in February that explain a lot of the licence clauses. OpenAI sent the House Select Committee on China a memo accusing DeepSeek of "free-riding" via obfuscated programmatic access. Then Anthropic published numbers: DeepSeek, Moonshot and MiniMax had run distillation campaigns through roughly 24,000 fraudulent accounts, with over 150,000, 3.4 million and 13 million exchanges respectively, targeting agentic coding and tool use specifically. Both labs tied this to export controls. The chip side has been chaotic: H200 exports to China were re-permitted in December with 25% of revenue to the US government, the BIS made them case-by-case in January with US-capacity and screening conditions, about ten Chinese firms were cleared, Beijing then restricted purchases, Commerce told Congress in July that "very few" had shipped, and NVIDIA took a $0.4 billion H200 charge in August and now excludes China data-centre compute from guidance entirely. The Chinese labs producing the models in my table are, in other words, training on hardware that is officially neither exported nor not exported, and shipping models that the US labs say are partly distilled from theirs. The revenue clauses in the licences are the Chinese labs' answer to the same question from the other side: if you are going to build a business on our weights, we want a contract.

And on 3 September NVIDIA announced it is acquiring Hugging Face for $12.93 billion, with the promise that the hub "will remain an open platform" and that "NVIDIA compute will not be required." The company that sells the picks now owns the store where the open models are kept. I take the promise at face value for now; I also note that the API calls above are the kind of thing that stops working when a platform's incentives change, and that a mirror is cheap.

What the EU actually requires now

This is the part I most wanted to get right, because the coverage is a mess. The AI Act's obligations on general-purpose AI model providers, Articles 53 and 55, applied from 2 August 2025. The Digital Omnibus on AI, proposed in November 2025 and in force as Regulation 2026/1744 since 27 July, delayed the high-risk system obligations, from 2 August 2026 to 2 December 2027 for the Annex III use cases and to 2 August 2028 for products under Annex I. It did not delay the GPAI obligations. And the AI Office's enforcement powers over GPAI providers became active on 2 August 2026: information requests, mandatory model evaluations, corrective measures, withdrawal from the market, and fines up to 3% of global turnover or €15 million. The Commission says it will start with "technical compliance dialogues" rather than fines. It also has the power not to.

What the GPAI obligations require of an open-weight provider, from the Commission's July 2025 guidelines: a model trained with more than 10²³ FLOP is presumed to be general-purpose; one over 10²⁵ FLOP is presumed to carry systemic risk and must be notified to the AI Office within two weeks, with the systemic-risk duties of evaluation, incident reporting and cybersecurity. The open-source exemption removes the technical documentation and downstream-information duties, but not the copyright policy and not the public summary of training data, and it does not apply at all above the systemic-risk threshold. Models placed on the market before 2 August 2025 have until 2 August 2027. Every model in my table over 700B is over 10²⁵ FLOP, and being MIT-licensed does not exempt any of them.

The Code of Practice has 21 signatories as of the 31 July list: Amazon, Anthropic, Google, IBM, Microsoft, Mistral, OpenAI, Cohere, Aleph Alpha, Black Forest Labs and others. xAI signed only the safety and security chapter. Meta refused in July 2025, citing "legal uncertainties", and has not signed since; the Chinese labs are not on the list either. Signing is voluntary and gives a presumption of conformity; not signing means the AI Office assesses you directly. For a European deployer of GLM-5.3 the practical question is who the "provider" is when the lab has no EU presence, and the answer in the Act is that the importer or distributor inherits obligations. I am not a lawyer and would check that with one before offering a 700B Chinese model to EU customers as a managed service; I do have EU customers, so I am checking.

The American picture is the mirror image. The July 2025 AI Action Plan has a section titled "Encourage Open-Source and Open-Weight AI". The December executive order set up a DOJ task force to litigate against state AI laws, and its first target was Colorado, where the DOJ intervened in xAI's suit in April and enforcement was suspended within days. California's SB 53 survived, effective 1 January, requiring frontier developers to publish safety frameworks and report critical incidents with penalties up to $1 million. So a US open-weight provider faces no federal model-level obligation and a shrinking set of state ones, while the same model in Europe is a regulated product. That asymmetry, more than anything about the models, will decide where the next generation of open weights is trained.

Running them

For the reader who wants to actually serve one of these, the hardware arithmetic in 2026 is: anything up to about 400B total parameters at FP8 fits one 8×H100 node with 640 GB of HBM (Qwen3.5-397B, GLM-5.3-Flash at 320B, Mistral Large 3). The trillion-class models need 8×H200 at 141 GB each, or two H100 nodes, or a 4-bit format; DeepSeek V4-Pro is about 860 GB at FP8 and Kimi K3 ships natively in MXFP4 for exactly this reason. At $2 to $4 per GPU-hour rented, a node is $12,000 to $24,000 a month, and the break-even against the open-model API prices above is somewhere between 5 and 20 billion tokens a month, which is a real product's volume and not a lab's. Below that, use the API; above it, vLLM or SGLang with expert parallelism, as described in the inference post. On a workstation, the 27B and 30B releases from Qwen, Meta and NVIDIA in GGUF or MLX are the point of this whole year for most people: Hugging Face counts 39.6 million Qwen GGUF downloads a month, and Apple's Foundation Models framework at WWDC 2026 gained an MLX backend that runs hub models on-device.

My own decision for the quarter: DeepSeek V4-Flash and Qwen3.8-27B as the defaults, both under plain permissive licences; GLM-5.3-Flash as the large option because its MIT licence has no thresholds; Kimi K3 and the full GLM-5.3 and Qwen3.8-2.4T offered only as bring-your-own-weights, because I am not going to be the party that owes Moonshot a contract on a customer's behalf. And a saved copy of every licence file with its commit hash, because three of them changed this year and one changed hours before release.

Related: LLM Inference in 2026, Seats Are Dying: the 2026 SaaS Repricing and The Worms Learned to Use Your AI Agent.