TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Mistral says it has released the weights for an updated version of Mistral Large 4, its 1-trillion-parameter multimodal model. The company reports gains on coding and defensive cybersecurity evaluations, but its benchmark results are company-provided and do not establish independent performance across real-world deployments.
Mistral has released the weights for an updated version of Mistral Large 4, a 1-trillion-parameter multimodal model, making it available for developers and organizations to deploy and customize. The company says three weeks of additional reinforcement learning improved coding and defensive cybersecurity performance, while new reasoning modes give users options to balance response quality against inference costs.
Mistral describes Large 4, also called “Le Chonk,” as its flagship open-weight model. It says the weights released now belong to an improved version of the model that entered public preview earlier this month. The company says feedback from developers, early-access customers and partners informed further training, though it did not provide details about the feedback or the training process beyond the additional reinforcement learning.
In results published by Mistral, the model scored 19.7% on Cognition’s private FrontierCode 2.0 benchmark, which assesses end-to-end code quality and whether a repository maintainer would merge an agent’s change. Mistral also reports a 48% score on the Artificial Analysis Coding Agent Index, tied with the leading open-weight model in its comparison while using 47% fewer tokens per task on average. On Terminal-Bench 4.0, the company says the model scored 41%, up 13 percentage points from its public-preview result, with 30% fewer output tokens.
The release adds three reasoning modes: low for simple tasks, medium for everyday use and high, the default, for tasks needing more precision and thorough reasoning. Mistral says the modes let developers adjust the trade-off between performance and cost. The company also reports a 51% score on the Artificial Analysis Cyber Index, which evaluates vulnerability discovery, reproduction, validation and patching, and says the model refused 96.6% of malicious cyber requests in its tests.
More Control Over Model Deployment
Releasing model weights can let organizations run a model in their own environment, adapt it to specific needs and reduce reliance on a hosted service. Mistral presents Large 4 as a foundation for self-hosted and customizable deployment, including for organizations seeking greater control over where data is processed. The announcement does not specify the release’s license terms or the infrastructure needed to operate a model of this size, both of which can shape who can use it in practice.
The reported token-efficiency results also speak to a practical cost question for coding agents: models that use fewer tokens for a task may be less expensive to run, depending on deployment and pricing. Mistral’s figures are benchmark results, not a guarantee of lower costs for every workflow. The three reasoning modes give developers another way to manage that balance, though the announcement does not quantify costs for each mode.
The cybersecurity claims matter because the same capabilities used to find and patch vulnerabilities can be misused. Mistral says its guardrails distinguish defensive work from malicious requests, and reports a high refusal rate on the latter. Those results offer a snapshot of the company’s tests; they do not settle how reliably safeguards will work across different prompts, users or deployments.
Products Worth Considering
Apple AirPods Pro 3 Wireless Earbuds
Bose Lifestyle Ultra Wireless Home Speaker
Logitech MX Brio Ultra HD 4K Webcam, Graphite
As an affiliate, we earn on qualifying purchases.
From Preview to Weight Release
Mistral says Large 4 first became available in public preview earlier this month. The company released the weights for the updated version after three additional weeks of reinforcement learning, saying coding gains continued during that period. The announcement frames the release as an extension of the preview rather than a separate model family.
The company also says Large 4 was trained from scratch on European infrastructure and has a documented lineage from training through release. Mistral connects those points to its “sovereign AI” positioning: organizations can use open weights to maintain more control over their AI systems. These are statements from the company; the release material provided does not independently audit the training history or define what operational controls a deployment requires.
Mistral’s announcement compares Large 4 with several other open-weight models across coding and security evaluations. One of the cited coding measures, FrontierCode 2.0, is described as a private benchmark, while other figures come from named external evaluation indexes. Readers should keep those different test settings in mind: benchmark scores measure performance on particular tasks and do not by themselves predict results across all software projects or security environments.
“Today, we’re releasing the weights of this improved Mistral Large 4.”
— Mistral
License, Costs and Independent Testing
The announcement does not state the specific license terms governing the weights, including any restrictions on commercial use or redistribution. It also does not set out the hardware requirements, deployment costs or availability details needed to assess how accessible a 1-trillion-parameter model will be to individual developers and smaller organizations.
The performance figures and safety results cited here come from Mistral’s release material. The company names external indexes and a private benchmark, but the supplied announcement does not include enough methodology to independently assess every comparison. It is also unclear how the reported results translate to customer workloads, how the safeguards perform outside the tested prompts, and whether further evaluations will change the model’s standing.
Deployment and Evaluation Ahead
Developers and organizations can now examine the released weights under the terms that accompany the release and test the model in their own coding, reasoning and defensive-security workflows. Those tests should clarify practical questions the announcement does not answer, including hardware needs, operating costs, task performance and guardrail behavior.
Mistral says it will continue working with developers, enterprise customers, partners and cybersecurity vendors. The company has not announced a date for a further model update or an independent evaluation milestone. The next useful evidence will come from the release documentation and reproducible evaluations that show how Large 4 performs beyond the results in Mistral’s announcement.
Key Questions
What did Mistral release?
Mistral says it released the weights for an updated Mistral Large 4, its 1-trillion-parameter multimodal model, following a public preview earlier this month.
What coding results did Mistral report?
The company reports a 19.7% score on FrontierCode 2.0, a 48% score on the Artificial Analysis Coding Agent Index and a 41% score on Terminal-Bench 4.0. These results are tied to specific benchmarks and should not be read as a guarantee of performance on every coding task.
What are the three reasoning modes?
Mistral lists low, medium and high modes. It recommends low for simple tasks, medium for everyday work and high—the default—for harder tasks requiring greater precision and more thorough reasoning.
What does Mistral say about cybersecurity safeguards?
Mistral reports a 51% score on the Artificial Analysis Cyber Index and a 96.6% refusal rate on malicious cyber requests in its tests. These are company-reported results, and the announcement does not establish how safeguards perform in every deployment.
Can anyone use or redistribute the weights?
The announcement says the weights are being released but does not specify the license terms in the supplied material. Users should check the release documentation for rules on access, commercial use and redistribution.
Source: mistral.ai
