AnswersPublished25 sources
Mistral Large 4’s Open Weights: Coding Gains, Cyber Claims and Caveats
Mistral released Large 4’s open weights on October 29, reporting a Terminal Bench 4.0 increase from 28% to 41% after additional reinforcement learning. The multimodal model has 1.05 trillion total parameters; its license is a modified MIT license, so users should check the terms before offering it as a hosted service.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Mistral released Large 4’s open weights on October 29, reporting a Terminal Bench 4.0 increase from 28% to 41% after additional reinforcement learning.
- The multimodal model has 1.05 trillion total parameters; its license is a modified MIT license, so users should check the terms before offering it as a hosted service.
- Mistral also reports a 51% Cyber Index score and a 96.6% refusal rate for malicious cyber requests; neither figure guarantees the open weights cannot be misused.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What are the key details and implications of France’s Mistral releasing the open weights of its flagship Mistral Large 4 model, nicknamed “L. Article summary: Mistral released the open weights of Mistral Large 4, “Le Chonk,” on October 29, following its October 6 preview, and claims leading open-weight performance in coding and defensive cybersecurity.[9] The available evidenc. Topic tags: general, general web, user generated, documentation. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks,
Mistral released the open weights of Mistral Large 4, nicknamed “Le Chonk,” on October 29, after previewing the model on October 6. The company presents the release as a stronger open-weight option for coding and defensive cybersecurity. Its reported benchmark gains are notable, but they should be read as task-specific results—not a verdict that the model leads every workload.4
7
9
18
What changed after the preview?
Mistral says additional reinforcement learning lifted Large 4’s Terminal-Bench 4.0 score from 28% in the preview to 41% in the released version, a 13-percentage-point increase.21
31 The company’s reported Artificial Analysis Coding Agent Index score is 48%, tied with GLM-5.3; reporting also says Large 4 used 47% fewer tokens per task in that comparison.
31
35
Those figures describe specific evaluations. They do not establish performance across all programming work, or show that a lower token count automatically means lower total operating costs. The available reporting also does not provide a substantiated, like-for-like numerical comparison with Anthropic’s Mythos 5.1.
A trillion-parameter multimodal model
Mistral’s documentation describes Large 4 as a multimodal mixture-of-experts model with 1.05 trillion total parameters and a 1.6-billion-parameter vision encoder.17 The parameter count needs a little context: the Hugging Face model description lists 49 billion active parameters per token, or 52 billion when embeddings and the output layer are included.
20
Mistral’s preview announcement said the model was trained using 3,800 Nvidia Grace Blackwell GPUs in the company’s European data centers.6 Release reporting describes three reasoning settings—low, medium and high—giving users a choice of reasoning effort.
34
Open weights, with a license to read
The weights are available through Hugging Face, and the published license is called the Mistral Ecosystem Open-Weight License. Its text describes it as a modified version of the MIT license.18
32 “Open weights” therefore should not be taken to mean that every kind of redistribution or hosted service is automatically covered by standard MIT terms. Anyone building a commercial offering should review the license itself and confirm which conditions apply to their use.
32
36
The practical significance is that teams can obtain the model weights rather than relying solely on access through a provider’s API. That can make self-hosting or model modification possible, though the release alone does not establish what infrastructure a particular deployment will require.
Cybersecurity claims need careful interpretation
Release coverage reports a 51% score on the Artificial Analysis Cyber Index and a 96.6% average refusal rate for malicious cyber requests.31
33 These are reported evaluation results, not guarantees that the model will refuse every harmful request or that releasing its weights carries no misuse risk.
A defensive-security score and resistance to malicious prompts measure different things. The figures should be considered alongside the evaluation details and the needs of a real deployment; they do not, on their own, settle the risks of a widely available model.
What the release means—and what it does not prove
Large 4 gives organizations another European-developed open-weight option in a market that includes both open and closed models from other regions. Mistral has highlighted the model’s European training infrastructure and its performance against open-weight competitors.6
15
18 For users, the meaningful distinction is the availability of weights to inspect and deploy—not a general conclusion about model quality based on a company’s country of origin.
The strongest supported takeaway is narrower: Mistral reports meaningful coding benchmark gains between preview and release, and a competitive coding-agent score alongside lower token use in one comparison. Those results may make Large 4 worth testing on a team’s own tasks. They do not establish broad superiority, independently confirm every cybersecurity claim, or remove the need to assess the license and deployment risks.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Mistral Large 4’s Open Weights: Coding Gains, Cyber Claims and Caveats"?
Mistral released Large 4’s open weights on October 29, reporting a Terminal Bench 4.0 increase from 28% to 41% after additional reinforcement learning.
What are the key points to validate first?
Mistral released Large 4’s open weights on October 29, reporting a Terminal Bench 4.0 increase from 28% to 41% after additional reinforcement learning. The multimodal model has 1.05 trillion total parameters; its license is a modified MIT license, so users should check the terms before offering it as a hosted service.
What should I do next in practice?
Mistral also reports a 51% Cyber Index score and a 96.6% refusal rate for malicious cyber requests; neither figure guarantees the open weights cannot be misused.