Open-weight models are the ones anyone can download and run. For most of the last three years the largest of them came from Meta, and the assumption in Europe was that an American company would keep publishing them. April 2026 was the month that assumption ended. This briefing sets out what was released, where the weights people actually use now come from, what the licences say, and what the EU rules require. It is written on 30 April with the sources available then; the September rankings and prices are in the update.
Meta closes, DeepSeek opens
Meta's Llama 4, released in April 2025, was received badly and its two-trillion-parameter Behemoth never shipped; the leaderboard episode around it is its own briefing.[1] The company spent the rest of the year rebuilding, buying half of Scale AI, forming Meta Superintelligence Labs and losing its chief scientist. On 8 April 2026 the lab shipped its first model, Muse Spark, a multimodal reasoning model with tool use and multi-agent capabilities, and Meta's first frontier model without open weights.[2] It is a capable model, and it is closed.
Sixteen days later DeepSeek released V4: a Pro model with 1.6 trillion parameters and 49 billion active per token, a Flash model with 284 billion and 13 billion active, both with a million-token context, both under the MIT licence that has covered every DeepSeek release since R1.[3] Between them, Alibaba's Qwen3.5 in February, Z.ai's GLM-5 with 745 billion parameters on 11 February, and Moonshot's Kimi K2.6 on 20 April filled out a Chinese open lineup that refreshes every six to eight weeks.[4][5][6] The US open releases of the same period were Nvidia's Nemotron 3 family and Google's Gemma 4 on 2 April, capable models in the small and mid range rather than the frontier; OpenAI has shipped nothing open since its gpt-oss pair in August 2025.[7][8]
Where the weights come from
- 1.15bn
- Cumulative Hugging Face downloads of Chinese open models by March 2026, against 723m for US models and 163m for EU models [9]
- 17.1%
- Share of all Hugging Face downloads by Chinese developers in the year to August 2025, against 15.8% for US developers [10]
- 4%
- EU share of new fine-tuned derivatives on Hugging Face by February 2026, down from 58% in early 2024 [9]
The download data tells the same story as the release calendar, with a lag. The ATOM Project, an academic tracking effort, counted 2.04 billion cumulative downloads across the models it follows by March 2026. Chinese models went from 97 million to 1.15 billion over its study window; US models from 177 million to 723 million; EU models from 65 million to 163 million. China overtook the US in late July 2025 and the gap has widened every month since. The share of new derivative models built on Chinese bases went from 10% to 70%, and the EU's from 58% to 4%. Alibaba's Qwen family alone accounts for about a billion downloads. On OpenRouter, a routing service that publishes which models its customers call, Chinese models took 72.7% of open-model tokens by January 2026.[9] Stanford's China programme found the crossover a different way: over the year to August 2025 Chinese developers accounted for 17.1% of all Hugging Face downloads against 15.8% for Americans, and in September 2025 Qwen overtook Llama as the most-downloaded family.[10]
How far behind the closed frontier
Epoch AI's measure is the one to use over time. In November 2024 it put the open frontier about a year behind the closed one.[11] Its October 2025 study narrowed that to about three and a half months, or seven points on its capabilities index, with DeepSeek's R1 the best open model at the time.[12] The gap has held in that region since, which means the open labs are a quarter behind and the closed labs are still ahead; a business choosing between them is choosing between last quarter's frontier at a fraction of the price and this quarter's at full price.
"Open" is doing a lot of work
The word covers at least three arrangements, and the difference matters to anyone deploying commercially. A plain Apache-2.0 or MIT licence lets you run, modify and sell without conditions: that is DeepSeek, GLM-5, Gemma 4, Mistral Large 3 and gpt-oss.[3][5][8][13][7] A second group ships the weights with a threshold. Kimi K2.6's modified MIT licence requires visible attribution above 100 million monthly users or $20 million a month in revenue, and Llama 4 caps use above 700 million users and bars EU-domiciled companies from its multimodal features.[6][14] A third group is open in name only for a business: MiniMax's M2.7 moved to a non-commercial licence in March 2026 after two MIT releases, shortly after the company listed in Hong Kong.[15] None of this is unreasonable, and most Irish businesses are nowhere near the thresholds. It does mean the licence file is part of the evaluation.
Europe's position
Mistral's Large 3, a 675 billion parameter model under Apache-2.0 released in December, is the largest European open model and the only European entry at frontier scale.[13] On 24 April Cohere and Aleph Alpha announced a merger creating a $20bn transatlantic company explicitly aimed at reducing dependence on American providers.[16] The Swiss Apertus project and the EuroLLM consortium publish smaller models with open training data. The ATOM figures say what that adds up to: a continent whose developers build on Chinese weights and whose own contribute 4% of new derivatives. That is not a reason not to use the Chinese models, which are excellent and cheap. It is a reason to know what one is building on.
The EU AI Act and open weights
The Act's obligations on providers of general-purpose models have applied since 2 August 2025, with the Commission's enforcement powers arriving on 2 August 2026. The open-source exemption is real but narrower than the phrase suggests: a model under a free and open licence is relieved of the documentation duties it would otherwise owe to deployers and regulators, but never of the copyright-policy requirement or the training-data summary, and the exemption does not apply to models designated as carrying systemic risk.[17] Models with revenue or user thresholds are unlikely to meet the "free and open" wording, and the trillion-parameter models almost certainly exceed the compute threshold that presumes systemic risk. None of this lands on an Irish company deploying the model; the obligations sit with the provider. It bears on which providers will still be serving Europe in a year.
