Glonce

Mistral releases open weights of 1-trillion-parameter Mistral Large 4

modelsThe Glonce editorsHacker News🔥87Glonce rating87Impact80Novelty70Relevance100Surprise60Players85Credibility90

Mistral AI has released the open weights of Mistral Large 4, nicknamed Le Chonk, its flagship 1-trillion-parameter, natively multimodal model. Earlier this month the company put the model out in public preview. Since then it spent about three weeks on further reinforcement learning, and the weights released today are of that improved version. Mistral says the new model also reasons more efficiently, which lowers inference costs.

Mistral's headline claim is that Mistral Large 4 is the best open-weight model for coding and among the most capable open models overall. It also says the model was trained from scratch on European infrastructure, and it pitches it as a base for sovereign AI: organizations can inspect, customize and deploy it on their own hardware without relying on Mistral to serve it through an API.

On coding, Mistral reports the following. On Cognition's private FrontierCode 2.0 benchmark, which judges end-to-end code quality and whether a repository maintainer would merge the agent's change, the model leads with 19.7%. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task on average. On Terminal-Bench 4.0 it improved by 13 percentage points to 41% compared with the public preview, taking the lead among open-weight models while cutting output tokens by 30%. In Surge Code's blind pairwise evaluations, the model was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models evaluated. The model now supports three reasoning modes that let developers trade performance against cost; the announcement text does not list them.

On cybersecurity, the model scores 51% on the Artificial Analysis Cyber Index, which measures finding vulnerabilities in a codebase, reproducing and validating them, and patching them without breaking existing functionality. Mistral says it handles legitimate defensive tasks without unnecessary refusals, while refusing malicious cyber requests at an average rate of 96.6%, which Mistral calls the highest it measures among open models. The company also admits a risk. After watching cybersecurity-related use in live traffic, it saw the model refuse black-box attacks against live targets, but still sees a potential for uplift to conduct real attacks. It is modifying its guardrails and will keep updating them in the months to come. A specialized variant for penetration testing, vulnerability assessment and red-teaming is in preview for enterprise design partners such as Orange Cyberdefense, with access subject to vetting and contractual safeguards and testing limited to owned or explicitly authorized systems in controlled environments.

For general work, Mistral reports a Finch score (spreadsheet-heavy finance and accounting workflows) of 73.3%, second only to Kimi K3 among open models. On BrowseComp it reaches 84.5%, up from an SFT checkpoint that scored 75%, which Mistral says puts it above GLM-5.3 and DeepSeek Flash 4.1 1017. On the B3 Agent Security Benchmark it resists 97.1% of attacks versus 93.3% in Preview, which Mistral says is the top of the open-source frontier. Demos include an agent that triages vulnerability reports and writes patches in a sandbox, a live 3D map of roughly 19,000 satellites and debris pieces called Orbital Trace that the model built and debugged itself, and a quantum chemistry solver written from scratch with an interactive 3D molecule page.

Availability: the weights are on Hugging Face for deployment with vLLM or SGLang. The model is also in Mistral Studio, the Mistral API, OpenRouter, Vibe Code, NVIDIA-hosted endpoints (as an optimized NVIDIA NIM container) and Microsoft Foundry, where Microsoft describes itself as the first cloud provider to bring the model to customers. Customization runs through Mistral Forge. The license is the Mistral Ecosystem Open-Weight license, a modified MIT license: free to use, modify, fine-tune and commercialize, with one condition. Businesses that operate a Model-as-a-Service offering must get a separate commercial license from Mistral before using the model commercially. Pricing starts at $0.70 per million input tokens. A full technical report is promised as coming soon.

The announcement includes a statement from Zakaria Moursli, Executive Vice President, Operational Performance, Payments and Digital Services at La Banque Postale, who calls the release a major investment by Mistral AI and says technological sovereignty is a key prerequisite for maintaining customers' trust.

Key facts

  • Mistral AI released the weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter, natively multimodal model, after a public preview earlier the same month and about three weeks of extra reinforcement learning.
  • Mistral reports 19.7% on Cognition's private FrontierCode 2.0, 48% on the Artificial Analysis Coding Agent Index (effectively tied with GLM-5.3, with 47% fewer tokens per task) and 41% on Terminal-Bench 4.0, up 13 percentage points from the preview.
  • On cybersecurity it reports 51% on the Artificial Analysis Cyber Index and a 96.6% average refusal rate on malicious cyber requests; Mistral also says it sees potential for uplift to conduct real attacks and is tightening guardrails.
  • License is the Mistral Ecosystem Open-Weight license, a modified MIT license: free to use, modify, fine-tune and commercialize, except that Model-as-a-Service operators need a separate commercial license from Mistral.
  • Weights are on Hugging Face (vLLM or SGLang); the model is also on Mistral Studio, the Mistral API, OpenRouter, Microsoft Foundry and NVIDIA endpoints, with input priced at $0.70 per million tokens.

Why it matters

A 1-trillion-parameter model with open weights is a large release from a European lab, and Mistral frames it as a sovereign alternative: trained from scratch on European infrastructure and runnable on a customer's own hardware. Mistral claims it is the best open-weight model for coding. If its numbers hold, it would sit level with GLM-5.3 on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task, which matters because token use drives the cost of running coding agents at scale. It also moves up from its own preview: Terminal-Bench 4.0 rose 13 percentage points to 41% in about three weeks of added reinforcement learning, and Mistral says gains show no signs of saturation.

Who it affects

Developers and teams building coding agents, who get a new open-weight option and three reasoning modes to balance cost and quality. Security teams and vendors, who can use the model for vulnerability reproduction and patching; vetted enterprise design partners such as Orange Cyberdefense get a specialized variant. Organizations with sovereignty or data-control requirements, such as La Banque Postale, which Mistral quotes. Companies that run hosted model services: under the license, Model-as-a-Service operators need a separate commercial license from Mistral.

How to use it

Download the weights from Hugging Face and serve them with vLLM or SGLang, or use the model through Mistral Studio, the Mistral API or OpenRouter. It is also offered as an optimized NVIDIA NIM container on NVIDIA-hosted endpoints, through Microsoft Foundry (with private deployment options for Azure developers), and in Vibe Code, where you can hand it a feature to build or a bug to fix in your own repository. Mistral Forge lets enterprises tailor it to proprietary data and guardrails. Input is priced at $0.70 per million tokens; the output price is not covered in this account. The license is a modified MIT license: free to use, modify, fine-tune and commercialize, apart from the Model-as-a-Service condition. The model card and license are the place to start, and a full technical report is promised.

How solid is it

This is a first-party announcement, and every benchmark figure is Mistral's own reported result; no independent replication is mentioned. Some comparisons come from third parties, namely Cognition's private FrontierCode 2.0 benchmark, the Artificial Analysis indices and Surge Code's blind pairwise evaluations, but the numbers are presented by Mistral. Several details are still missing: the announcement does not name the three reasoning modes, and it gives no active-parameter count, architecture, context window or training data details. The technical report has no stated release date beyond coming soon. The cybersecurity vendor quotes the page promises are absent from this account.

Risks and caveats

Mistral itself flags a dual-use risk: it observed the model refusing black-box attacks against live targets in live traffic, yet sees a potential for uplift to conduct real attacks, and says it is modifying its guardrails and will keep updating them. The 96.6% refusal rate is an average and a Mistral measurement. The specialized offensive-security variant is limited to vetted partners under contractual safeguards. The license is not plain MIT: Model-as-a-Service businesses need a separate commercial license. No hardware requirements for self-hosting a 1-trillion-parameter model are given.

“Organizations have the ability to inspect, customize, and deploy the model on their own infrastructure.”

— Mistral AI, announcement of Mistral Large 4