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 AI says it has released weights for an improved version of Mistral Large 4, its 1-trillion-parameter multimodal model, after three weeks of additional reinforcement learning following a public preview. The company reports stronger coding benchmark results, greater token efficiency and a 96.6% refusal rate on a set of malicious cyber prompts; independent verification and full release terms are not detailed in the supplied announcement.
Mistral AI says it has released the weights for an updated version of Mistral Large 4, a 1-trillion-parameter, natively multimodal model, following three weeks of additional reinforcement learning after its public preview. The release gives developers and organizations access to model weights they can deploy and customize, while Mistral says the update improves coding performance and uses fewer tokens on some coding tasks.
Mistral describes Large 4, also called “Le Chonk,” as its flagship model and says the new weights reflect feedback from developers, early-access customers and partners. The company says the model was trained from scratch on European infrastructure. The announcement does not specify the license, hardware requirements, or all conditions governing use and redistribution, so “open-weight” should not be taken to mean those terms are unrestricted.
The company reports a 19.7% score on Cognition’s private FrontierCode 2.0 benchmark, which it says evaluates end-to-end code quality and whether a repository maintainer would accept an agent’s changes. Mistral also cites a 48% score on the Artificial Analysis Coding Agent Index, describing that result as effectively tied with the leading open-weight model while using 47% fewer tokens per task on average. These are reported benchmark results; the announcement does not provide the full evaluation setup or independent replication.
On Terminal-Bench 4.0, Mistral reports a score of 41%, 13 percentage points above its public-preview result, alongside a 30% reduction in output tokens. It says Large 4 now offers low, medium and high reasoning modes: low for simpler tasks, medium for everyday use, and high as the default for demanding tasks. The modes are intended to let users adjust the balance between reasoning effort and cost.
Lower Token Use for Coding Agents
The release matters to organizations testing AI coding agents because model weights can offer more control over where a model runs and how it is adapted than a hosted service alone. Mistral presents Large 4 as an option for teams seeking self-hosted deployment and customization, including institutions with requirements around data control. Whether those benefits are practical will depend on the release license, infrastructure costs and the model’s performance in each organization’s own workflows.
Token efficiency is a potentially meaningful part of the company’s pitch. Mistral says Large 4 matched a leading open-weight model on one coding index while using fewer tokens per task. If reproduced in real deployments, reduced token use could lower inference costs or allow more agent activity within a fixed budget. The announcement does not provide enough information to calculate total operating costs, which also depend on hardware, latency, task success and the amount of human review required.
Mistral also frames the model as suitable for defensive cybersecurity work. It reports a 51% Artificial Analysis Cyber Index score and an average refusal rate of 96.6% on malicious cyber prompts from three named prompt sets. That distinction matters because security teams need systems that can assist with finding and fixing vulnerabilities without facilitating harmful requests. The figures are company-reported and do not establish how the safeguards perform across all real-world security tasks.
Products Worth Considering
EcoFlow Smart Home Panel 2 Whole-Home Backup System Hub
DJI Mic 3 Wireless Microphone System (2 Transmitters, 1 Receiver, Charging Case)
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
From Preview to Released Weights
Mistral says Large 4 first entered public preview earlier this month. The weight release follows three weeks of additional reinforcement learning, which the company says improved coding and defensive cybersecurity performance. The announcement presents the release as a response to feedback from developers and early-access users, but does not name the feedback that led to specific changes.
The model’s scale and multimodal design are central to Mistral’s description: it has 1 trillion parameters and is described as natively multimodal. Mistral says it was trained from scratch on European infrastructure and emphasizes a clear lineage from training through release. Those statements support the company’s case for sovereign AI deployments, but the announcement does not spell out data residency, ownership, or audit provisions for particular customers.
The cited evidence combines different kinds of evaluation. FrontierCode 2.0 is identified as a private benchmark; the Coding Agent Index and Cyber Index are attributed to Artificial Analysis; and Surge Code is described as a blind human-preference evaluation. Results from these tests are not directly interchangeable. Mistral also says Surge Code ranked its model behind only Claude Opus 5.5 among those evaluated, but the supplied material does not give the full test sample or evaluation protocol.
“Today, we’re releasing the weights of this improved Mistral Large 4.”
— Mistral AI
Release Terms and Results to Verify
The announcement does not state the precise license, download location, or release date, and it does not detail hardware requirements or whether all model components are included with the weights. Those details affect whether and how developers can run, modify and redistribute the model.
Most performance figures come from Mistral’s account of evaluations. Some tests are attributed to outside organizations, but the supplied announcement does not include complete methodology, sample sizes, confidence intervals or independent replication. It is also not clear how the three reasoning modes compare in latency and total cost across real applications, or how the cybersecurity refusal rate generalizes beyond the named prompt sets.
Deployment and Independent Testing
The immediate next step for developers is to review the release materials for the license, model files and deployment requirements, then test Large 4 against their own coding, reasoning and security workloads. Independent evaluations and reports from organizations running the weights could clarify how the company’s benchmark results translate to real tasks, including token use, reliability and total cost.
Mistral says it will continue scaling reinforcement learning and describes performance gains as ongoing, but it gives no specific date for another update. Further detail about the model’s safeguards, release terms and supported deployment options will be needed before organizations can judge how well it fits their requirements.
Key Questions
What did Mistral release?
Mistral says it released weights for an updated version of Mistral Large 4, a 1-trillion-parameter, natively multimodal model that had previously been in public preview.
What does open-weight mean here?
It means the model weights are being made available, allowing developers to run or adapt the model subject to the release terms. Mistral’s supplied announcement does not state the license or redistribution conditions, so those should be checked in the release materials.
What performance gains does Mistral report?
Mistral reports a 19.7% FrontierCode 2.0 score, a 48% Coding Agent Index score with 47% fewer tokens per task on average than the comparison it describes, and 41% on Terminal-Bench 4.0. These are reported evaluation results, not a guarantee of performance on every task.
What cybersecurity safeguards does the company cite?
Mistral reports a 96.6% average refusal rate on malicious cyber prompts from JailbreakBench, StrongREJECT and AgentHarm, alongside a 51% score on the Artificial Analysis Cyber Index. The announcement does not establish how those figures generalize to all security situations.
When can developers use the weights?
Mistral says the weights are being released, but the supplied announcement does not provide a direct download link, exact release date or hardware requirements. Developers should consult Mistral’s release materials for current access and license details.
Source: mistral.ai
