On October 29 Mistral put the weights of Mistral Large 4 on Hugging Face, three weeks after an API preview on October 6. The model is a 1.05-trillion-parameter mixture of experts with 52 billion parameters active per token, released under what Mistral calls "a modified MIT license": free to run, tune and sell, unless your business is selling the model itself as a service. It closes a month in which Meta, Reflection and Mistral all published open weights above 500 billion parameters, and in which licenses, not parameter counts, decide who can serve them.
A trillion parameters at half the list price
Mistral's docs add a 1.6-billion-parameter vision encoder and a 1-million-token context, and a developer's reading of the model card counts 49 billion active parameters in the experts and 52 billion with embeddings. Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, and says it is fluent in more than 160 languages, every official EU language among them.
The release is not the preview. Three more weeks of reinforcement learning lifted Terminal-Bench 4.0 to 41%, 13 points above the preview, on 30% fewer output tokens. Mistral's own figures put it at 48% on Artificial Analysis's Coding Agent Index, level with GLM-5.3 on 47% fewer tokens per task. The independent Intelligence Index scored the preview 38, tied with GPT-6 Luna and the highest of any model from outside the US and China, behind three cheaper Chinese open-weights models: MiMo-V2.6-Pro at 46, GLM-5.3-Flash at 42 and DeepSeek V4.1 Flash at 39.
The price is the quiet headline. The open-weights post sets the API at $0.70 per million input tokens, $0.07 cached and $2.10 output, about half the $1.36, $0.14 and $4.18 list price in Mistral's model docs, which the October 6 changelog paired with a two-week 50% launch discount. The discount, in effect, became the price; the docs still showed the old list when we checked.
Three open giants in a fortnight
Large 4 did not arrive alone. Meta published Muse Spark Open, 816 billion parameters with 42 billion active, on October 15. Reflection, which announced Beam, 501 billion parameters with 23 billion active, on October 5, released its weights under Apache 2.0 on October 16. Moonshot has promised weights for Kimi K3.1 Pro, a post-trained Kimi K3, by November 4; K3 has 2.8 trillion parameters.
Strip away the totals and the three released models look like siblings: each wakes about 5% of its weights for any given token (our arithmetic). On the Intelligence Index, Muse Spark Open scores 39 and Beam 30, which puts Large 4's preview a point behind America's best open model and well clear of Beam.
The license decides who can serve it
Here the siblings part ways. Beam's model card reads simply apache-2.0, and Apache and MIT carry no revenue thresholds or naming requirements. Anyone may host Beam, resell it or fold it into a cloud without asking.
Meta's license lets anyone run Spark Open as a service until group revenue passes $20 million in a 12-month period, after which they must notify Meta and sign a separate agreement, and every product must display a "Built with Muse Spark Open" label.
Mistral's license has no threshold and asks for no label. It has one condition: a business operating a model-as-a-service offering needs a separate commercial license from Mistral before commercial use. The license defines that business as giving third parties access in which they can "exercise meaningful control over the inputs, parameters, or training data", with carve-outs for capabilities embedded in end-user products and for relaying requests to models hosted elsewhere, per a developer's reading of the license file.
So "can I use it?" depends on who asks. An enterprise running it internally, or a software company embedding it in a feature, owes nobody anything under any of the three. A small inference startup reselling tokens is free under Beam, free under Meta until its revenue passes $20 million, and needs a contract with Mistral on day one. A hyperscaler is free only under Beam.
Mistral, which sells its own API at $0.70 and $2.10, has written a license that protects exactly that business. The Batch lists Beam at the same $0.70 and $2.10 per million tokens, as if the going rate for a very large open model had already settled.
What 3,800 GPUs buy
For scale, Reflection pretrained a model half Large 4's size on 6,144 GB300 NVL72 GPUs in under four weeks, then ran reinforcement learning on 10,500 for four more. Meta says it trained Muse Wonder at Prometheus, a gigawatt-scale data center; Mistral's reinforcement-learning stage drew about 10 megawatts in France on roughly 3,000 GPUs. The figures measure different things, but the orders of magnitude are the story.
Read charitably, that is efficiency: the best model outside the US and China, on a smaller fleet than its American peers. Read strictly, the three Chinese models ahead of it on the index are also cheaper, and Mistral's announcements do not say how long training took, so compute-per-point comparisons are guesswork. The sovereignty case survives both readings. Pierre Stock, Mistral's VP of science, argues no company should depend on a provider that can cut off access "from one day to another". A weights file on your own disks cannot be revoked.
Two caveats belong here. The public release is not the most capable version: a variant with limited safeguards and extra cybersecurity capabilities is reserved for selected companies, partners and government agencies, and Mistral reports a 96.6% refusal rate on malicious cyber requests for the open one. And self-hosting is not casual: The Next Web reports a quantized build runs on four Nvidia B200 or B300 GPUs, and at four bits the weights alone come to about 525 GB (our arithmetic), against about 300 GB for Beam.
The 30,000-foot read
Parameters have become the cheap part. October produced three open models above 500 billion parameters with nearly identical sparsity. What still differs is permission. Reflection gives the weights away and plans to sell the deployment around them; Meta gates the big hosts; Mistral gates every host. Each license is a business model written in legal prose, and each is candid about what it protects.
For Europe, the meaningful fact is not that Large 4 trails three Chinese models on one index. It is that a European company trained a trillion-parameter model on its own soil and its own GPUs, then handed out a copy. Sovereignty is not leading the leaderboard; it is owning something nobody can switch off. Watch for Kimi K3.1 Pro's weights and license on November 4, Artificial Analysis's score for the released Large 4, and whether any hyperscaler signs Mistral's commercial license.
Sources: Mistral, "Mistral Large 4 (Le Chonk) is now open weight" (primary: license name and condition, API price, Mistral-reported benchmarks, languages, three weeks of extra reinforcement learning, cyber refusal rate, Hugging Face repo name); Mistral, "Introducing Mistral Large 4" (primary: October 6 preview, 3,800 Grace Blackwell GPUs, roughly 3,000-GPU reinforcement-learning setup); Mistral docs, Mistral Large 4 (primary: 1.05T total, 52B active, 1.6B vision encoder, 1M context, list price); Mistral changelog (primary: two-week 50% launch discount); The Next Web (October 29 weights release, 10 MW reinforcement learning in France, four-GPU quantized build, Stock remark); The Batch, issue 375 (Intelligence Index score and peers, EU languages, restricted-safeguards variant); The Batch, issue 376 (index scores for Muse Spark Open and Beam, Beam's per-token price and 4-bit footprint); DEV Community, Gabriel Anhaia (license clause wording, 49B and 52B split); Muse Spark Open model card (816B and 42B, Muse Model License terms); Beam model card (Apache 2.0 license line); Reflection, "Introducing Beam" (October 5 announcement, 501B and 23B, GB300 counts); Unite.AI (October 16 Beam weights release); VKTR on Kimi K3.1 Pro (November 4 weights promise, K3 size); VKTR on Beam (Reflection's deployment business model); The Decoder (Meta's Prometheus training claim); Trending Topics (Apache and MIT carry no thresholds or naming rules). CAVEATS and corrections: Correction: early notes priced the API at $1.36, $0.14 and $4.18 per million tokens; that is the preview list price in Mistral's docs, while Mistral's open-weights announcement sets $0.70, $0.07 and $2.10 from launch day, and the docs still showed the old list when checked. Correction: Mistral says 3,800 Grace Blackwell GPUs, not about 4,000, and the reported two-month training run could not be confirmed on Mistral's pages. The 49B (experts) versus 52B (with embeddings) split comes from a developer's reading of the model card; Mistral's docs say 52B active. We could not open the Hugging Face model card or license file; the license's core condition is confirmed on Mistral's announcement, and clause-level wording comes from that developer write-up. Mistral names the repo mistralai/Mistral-Large-4-1T-A52B, which differs from the URL in early notes. The Intelligence Index score of 38 was measured on the API preview, before three more weeks of reinforcement learning, and comes from an earlier issue of The Batch than the Muse Spark Open and Beam scores. Terminal-Bench, Coding Agent Index and cyber refusal figures are Mistral-reported; the Prometheus claim is Meta's. Beam's October 16 release date comes from Unite.AI, which labels the piece as produced by an AI research agent with editorial review; Beam's specs and license were confirmed on Reflection's own pages. Kimi K3.1 Pro details are from VKTR; its license is not yet public. The GPU and power comparisons mix hardware generations and training stages and are indicative only; we found no sourced figure for Chinese labs' training clusters and make no claim about them. The 5% active share and the 525 GB figure are our arithmetic.