Key points
- Mistral Large 4 leads the FrontierCode benchmark with a score of 19.7 percent.
- The model uses 47 percent fewer tokens than GLM-5.3 on the Artificial Analysis index.
- Weights are now openly available for self-hosted deployment after a public preview.
Mistral AI released the open weights for Mistral Large 4, a one-trillion-parameter multimodal model, according to an official advisory from the company.reported The release follows a public preview that began earlier this month, with the company citing improved coding and cybersecurity capabilities. The model is described as the best open-weight option for coding tasks, trained entirely on European infrastructure.
How we got here
Mistral entered the public preview phase with strong initial results on leading coding benchmarks, according to the company's advisory.company claim Feedback from developers and early-access customers was collected during this period. The company stated that it used this feedback to scale reinforcement learning, aiming to drive improvements in coding and defensive cybersecurity before the full release.
What is new
The newly released weights show substantial gains in reasoning efficiency, lowering inference costs while maintaining performance, according to Mistral.reported On Cognition's private FrontierCode 2.0 benchmark, the model scored 19.7 percent, leading open-weight competitors.reported It also improved by 13 percentage points on Terminal-Bench 4.0 to reach 41 percent, while reducing output tokens by 30 percent compared to its preview version.
Reading between the lines
Mistral Large 4 supports three reasoning modes to help developers balance cost and performance for different use cases, according to the advisory.reported The company noted that gains continue to accelerate with no signs of saturation. It also partnered with leading cybersecurity vendors to evaluate the model in real-world environments, testing its strengths in vulnerability reproduction and patching capabilities.
The open questions
The advisory claims the model leads open-weight benchmarks, but these results come from specific tests like FrontierCode and Terminal-Bench, which may not reflect all coding scenarios. Human preference tests by Surge Code placed it behind Claude Opus 5.5, suggesting limits in general utility despite coding strength.reported The company states it is working with vendors on real-world cybersecurity tests, but no independent results from those evaluations are provided in the material, leaving the practical defensive security claims unverified by third parties.
Why it matters
- Evaluators can now access the weights to independently verify the claimed 19.7 percent score on the FrontierCode 2.0 benchmark.
- The 47 percent reduction in token usage compared to GLM-5.3 offers a concrete metric for cost-efficiency analysis in agentic coding workflows.
- The availability of three reasoning modes allows testers to assess performance trade-offs between precision and inference speed for specific tasks.
- The lack of independent cybersecurity validation means evaluators must treat the defensive security claims as preliminary until third-party results emerge.
Topics: Mistral AI, open weights, coding benchmarks, multimodal models, reinforcement learning
Updated
Numbers in context
Circle area is proportional to the magnitude; labels retain signs. Figures may describe different measures. Read their contexts above.Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
What is the score of Mistral Large 4 on the FrontierCode 2.0 benchmark?
Mistral Large 4 scored 19.7 percent on Cognition's private FrontierCode 2.0 benchmark, leading among open-weight models.
How does Mistral Large 4 compare to GLM-5.3 in terms of token usage?
On the Artificial Analysis Coding Agent Index, Mistral Large 4 uses 47 percent fewer tokens per task on average than GLM-5.3.
Are the weights for Mistral Large 4 available for self-hosted deployment?
Yes, Mistral AI released the weights openly, allowing for customizable and self-hosted deployment options.
Reading receipt
- Mistral Large 4 leads the FrontierCode benchmark with a score of 19.7 percent.
- The model uses 47 percent fewer tokens than GLM-5.3 on the Artificial Analysis index.
- Weights are now openly available for self-hosted deployment after a public preview.
