Mistral Previews 1-Trillion-Parameter Large 4, Pitching Open Weights Against Lock-In

Earlier this year, Hugging Face was breached by rogue OpenAI agents. It turned to leading American closed models to help defend itself, only to find their guardrails blocking the work. The guardrails mistook Hugging Face’s defensive actions for an attack.

Hugging Face ended up using Z.ai’s GLM 5.2, a Chinese open-weight model.

Mistral on Tuesday launched a public preview of Mistral Large 4, nicknamed Le Chonk, with weights due Oct. 27. The French company is selling open weights as protection against a closed vendor turning a capability off.

The preview

Guillaume Lample, Mistral’s co-founder and chief scientist, is a former Meta researcher. Mistral says its science team has grown from three researchers to roughly 300. Hugging Face’s blocked defensive work is the kind of limitation Lample describes.

On cyberdefense, Lample said, “you cannot afford to be vulnerable to the fact that the model you are using to protect yourself might disappear one day, or might be too limited.”

Mistral reports Large 4 among the top five models on the Artificial Analysis Cyber Index, whose public evaluations do not yet list it. On a test requiring models to reproduce and patch a vulnerability, Mistral reports an 82% score, the highest of any model. Mistral says several leading closed models score near zero because they refuse the task. It also reports 93% on Cybench.

Until the weights release, cybersecurity leaders, vetted partners and state authorities are testing the model with reduced moderation and expanded cyber capabilities.

Open weights means the trained parameters are meant to be downloadable, so a customer can hold a complete copy that no vendor can retire. Each model version behaves slightly differently, so a forced migration to a newer model can break systems built around the old one.

“But what really matters is to own the model, even for US companies,” Lample said. “If you use a closed model, there is no guarantee it will still be there tomorrow.”

Large 4 has about a trillion parameters in total, with 49 billion active during inference. The preview accepts multimodal input and produces text. It supports more than 160 languages.

Mistral trained it from scratch over about two months in its European data centers, using Nvidia Grace Blackwell GPUs, with the launch post giving 3,800 GPUs and reports citing Mistral putting it at about 4,000. Nvidia CEO Jensen Huang said on X that OpenAI trained GPT-6 Astra on roughly 100,000 GPUs. Those figures show the difference in fleet size, but competitors do not publish training compute on a uniform basis, so the gap is not an exact ratio of training compute.

Its previous Large 3 had 675 billion total parameters, with 41 billion active. That model was trained on 3,000 Nvidia H200 GPUs.

The US government accused Chinese labs of abusing distillation, training smaller models on larger models’ outputs. Mistral says it trained from scratch. “We are fully separate from other models, and we don’t take inspiration from them,” Pierre Stock, Mistral’s VP of Science, said.

The benchmarks

Mistral’s launch post gives Large 4 a 61.7% DeepSWE v1.1 score, rounded to about 62% in its chart. That chart puts Reflection AI’s Beam at 44% and Qwen 3.8 Max at 51%. GLM-5.3 sits close to Mistral’s result.

But the live DeepSWE leaderboard, selecting each model’s best published configuration, puts GLM-5.3 and Kimi K3 at about 69% as of Oct. 6. Closed models from OpenAI, Google and Anthropic sit around 74% on that board. In a blind coding evaluation Mistral ran with Surge AI, Large 4 ranked second among five models, scoring 3.74 behind a Claude model at 4.22 and ahead of Kimi K3 at 3.59.

Large 4 is absent from the DeepSWE board. Its ranking claim is provisional until outsiders can test the final model and weights.

Andrea Renda, director of research at the Centre for European Policy Studies, assessed Mistral’s position in July, before this launch. EU sovereignty ambitions and American hostility made “a magic formula that all of a sudden puts Mistral, whose performance has not been spectacular, in a favorable position,” Renda said.

On Harvey’s legal benchmark, Mistral’s chart gives Large 4 a 15% task-pass rate. The public Vals.ai board puts Kimi K3 at 12.92%, MiMo V2.6 Pro at 10.83% and GLM-5.3 at 8.33%. That comparison holds if Mistral used the same method as the public board.

The limits

The Trump administration temporarily restricted distribution of OpenAI and Anthropic models in June over cyberattack concerns. The reminder that the US government could revoke access opened a door for a Europe-based open-weight lab. The White House has reportedly asked American labs to withhold unreleased models even from the UK’s AI Safety Institute.

Mistral raised €3 billion in a Samsung-led Series D on Sept. 8, at a post-money valuation above €21 billion. Mistral says it supports more than 125 enterprises, including Airbus, ASML and HSBC. CEO Arthur Mensch, a former DeepMind researcher, expects annual recurring revenue above $1 billion this year.

Large 4 is too large to run on a desktop or laptop and may be impractical even for some universities. Until Oct. 27, the preview is served on Mistral’s own infrastructure.

Its previous Large 3 used Apache 2.0. Large 4’s weights are due Oct. 27 under an expected custom Mistral license. As of launch, Mistral’s Hugging Face page lists no Large 4 checkpoint. The planned release would let outside evaluators test the final weights.

Moonshot Releases Kimi K3 Weights as Amodei Rejects Open-Weight Ban

Moonshot AI released the weights for its Kimi K3 model on Monday, a system with 2.8 trillion parameters. Developers can now download, modify and self-host K3, which the report described as the world's

The Implicator

Meta Releases 30B Open-Weight Muse Glimmer and Promises Spark 1.2 Weights

Meta released Muse Glimmer, a 30-billion-parameter open-weight model, Monday. Glimmer distills Muse Spark to run local agents on a single high-end Mac or PC. Meta promised Muse Spark 1.2 weights in co

The Implicator

Alibaba Publishes the Qwen3.8-Max Benchmarks It Withheld Two Weeks Ago

Alibaba on Monday published benchmark scores it had withheld on July 19 for Qwen3.8-Max, its 2.4 trillion-parameter model. The launch post says Qwen3.8-Max has 95 billion active parameters. Many outle

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai