Readers get a fully inspectable model they can keep, as Mistral launched a public preview of Mistral Large 4 (nicknamed Le Chonk) on Oct. 6, 2026, with the model weights scheduled for release on Oct. 27. The French AI lab is framing downloadable weights as a structural safeguard: once a customer holds a complete copy of the trained parameters, no vendor can quietly retire the system that depends on them.
What is in the Mistral Large 4 preview?
Mistral Large 4 is a multimodal model that accepts multiple input types and produces text. It carries roughly one trillion parameters in total, with about 49 billion active during inference. The previous Mistral Large 3 had 675 billion total parameters with 41 billion active. Large 4 supports more than 160 languages and was trained from scratch over about two months in Mistral’s European data centers on a fleet of Nvidia Grace Blackwell GPUs. The launch post cites 3,800 GPUs, while outside reports put the figure closer to 4,000. Mistral says its science team has grown from three researchers to roughly 300.
Until the weights release on Oct. 27, the preview is served on Mistral’s own infrastructure. Cybersecurity leaders, vetted partners, and state authorities are testing the model with reduced moderation and expanded cyber capabilities. The weights are expected under a custom Mistral license, and the model is too large to run on a desktop or laptop and may be impractical even for some universities. As of launch, Mistral’s Hugging Face page lists no Large 4 checkpoint.
How does Mistral frame open weights as anti-lock-in?
Mistral’s argument turns on a specific risk: a closed model can disappear, be restricted, or refuse a legitimate task the customer needs. Mistral points to a recent incident in which a platform under attack turned to leading American closed models for defensive help, only to have their guardrails mistake the defensive actions for an attack. The platform ultimately used a Chinese open-weight model to continue its defense.
The lab also points to a separate moment earlier this year, when a US government action temporarily restricted distribution of certain American models over cyberattack concerns. Mistral argues that holding the weights means holding a complete copy that no vendor can retire, so a defensive mission shifts every time the model is forced to change. The framing is that European open-weight infrastructure gives enterprises and government buyers a fallback that does not depend on any single vendor’s policy choices.
How does Mistral Large 4 score on cyber and coding benchmarks?
Mistral reports an 82% score on a cyber test that requires models to reproduce and patch a vulnerability, the highest of any model, while several leading closed models score near zero because they refuse the task. The company also reports 93% on Cybench, a public cybersecurity evaluation. Mistral says Large 4 ranks among the top five models on the Artificial Analysis Cyber Index, though that index’s public evaluations do not yet list it.
On coding, Mistral’s launch post gives Large 4 a 61.7% DeepSWE v1.1 score, rounded to about 62% on its chart. That chart places Reflection AI’s Beam at 44% and Qwen 3.8 Max at 51%, with GLM-5.3 sitting close to Mistral’s result. However, the live DeepSWE leaderboard, using each model’s best published configuration, puts GLM-5.3 and Kimi K3 at about 69% as of Oct. 6, with closed models from OpenAI, Google, and Anthropic around 74%. Large 4 is not yet 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. On Harvey’s legal benchmark, Mistral’s chart gives Large 4 a 15% task-pass rate, compared with Kimi K3 at 12.92%, MiMo V2.6 Pro at 10.83%, and GLM-5.3 at 8.33% on the public Vals.ai board, if Mistral used the same method. Mistral’s ranking claim is provisional until outside evaluators can test the final weights.
Where does Mistral stand commercially?
Mistral raised €3 billion in a Samsung-led Series D on Sept. 8, 2026, at a post-money valuation above €21 billion. The company says it works with more than 125 enterprises, including Airbus, ASML, and HSBC, and it expects annual recurring revenue above $1 billion this year. The lab also points to its training approach as a point of separation. The US government has accused Chinese labs of abusing distillation, training smaller models on the outputs of larger models. Mistral says it trained Large 4 from scratch and did not draw on other models’ outputs.
How does this fit the wider open-weight picture?
Independent research backs the broader case that open-weight models can match closed-weight systems on specific tasks. A 2025 study published in Radiology (Nowak et al., DOI: 10.1148/radiol.240895) compared 17 open-weight LLMs with OpenAI’s GPT-4o on extracting findings from chest radiography reports. The highest zero-shot macro-averaged F1 score on English reports came from GPT-4o at 92.4%, but Mistral-Large reached 92.6%, Llama-3.1-70b reached 92.2%, and Llama-3.1-405b reached 90.3%, all comparable to GPT-4o and well above the rule-based CheXpert baseline at 73.1%. On German reports, Mistral-Large scored 91.6%, outperforming rule-based annotation at 74.8%. A second 2025 Radiology study (DOI: 10.1148/radiol.241051) found that fine-tuned Mistral-7B showed substantial agreement with radiologists (mean κ 0.73 to 0.77) and rivaled GPT-4 Turbo in weighted overall mean accuracy (91% vs 92%, P = .31) on extracting clinical history elements. Together these findings support the broader argument that open-weight models can be competitive with closed-weight systems on defined tasks, including in regulated settings where local deployment matters.
FAQ
What is Mistral Large 4?
Mistral Large 4, nicknamed Le Chonk, is a multimodal model with about one trillion parameters in total and 49 billion active during inference. Mistral launched a public preview on Oct. 6, 2026.
When will Mistral Large 4 weights be released?
Mistral says the weights are due Oct. 27, expected under a custom Mistral license. Until then, the preview is served on Mistral’s own infrastructure.
Can Mistral Large 4 run on a laptop?
No. The model is too large to run on a desktop or laptop and may be impractical even for some universities.