Mistral open-weights Mistral Large 4, its 1-trillion-parameter model for coding and defense

Mistral has published the weights of Mistral Large 4, a 1-trillion-parameter, natively multimodal model that the company describes as its flagship system. The release, announced on Mistral's site and surfaced via Hacker News, follows a public preview earlier in October. According to Mistral, about three weeks of additional reinforcement learning since that preview drove large gains in agentic coding and defensive cybersecurity, and the company says the improvements are still accelerating.

Coding results lead the open-weight field

Mistral claims the model is now the strongest open-weight option for code generation. On FrontierCode 2.0, a private benchmark from Cognition that grades end-to-end change quality — specifically whether a repository maintainer would accept an agent's patch — Mistral Large 4 leads with 19.7%, narrowly ahead of DeepSeek V4.1 Flash at 19.3% and GLM-5.3 at 17.8%.

On the Artificial Analysis Coding Agent Index, the model scores 48%, which Mistral characterises as effectively tied with GLM-5.3, the leading open-weight entry, while consuming 47% fewer tokens per task on average. Terminal-based work improved too: Terminal-Bench 4.0 climbed 13 percentage points from the preview to 41%, ahead of other open-weight models while cutting output tokens by 30%.

Three reasoning modes for cost control

To let developers trade cost against accuracy, the release ships with three reasoning settings: Low for simple jobs, Medium for everyday work, and High, the default, for tasks that demand careful, thorough reasoning. Mistral frames the improved reasoning efficiency as a way to keep capable coding agents affordable to operate at scale.

Human evaluators agree, mostly

The benchmark numbers are paired with preference testing. In blind pairwise comparisons run by Surge Code, Mistral Large 4 was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking behind only Anthropic's Claude Opus 5.5 among the models evaluated.

A security-first pitch, with one caveat

The release leans heavily into defensive cybersecurity. Mistral Large 4 scores 51% on the Artificial Analysis Cyber Index, which tests whether a model can find, reproduce, validate and patch vulnerabilities in a codebase without breaking existing functionality — and, notably, whether it will do that work without refusing legitimate defensive requests.

On the safety side, Mistral reports an average refusal rate of 96.6% for malicious cyber requests, drawn from cyber-related prompts in JailbreakBench, StrongREJECT and AgentHarm, which it calls the highest among the open models it measured.

One tension is worth flagging: although Mistral's summary claims leadership in defensive security, the company's own comparison chart places DeepSeek-V4.1-Flash 1017 at 57% and MiMo-V2.6-Pro at 56%, both above Mistral Large 4's 51% on the Cyber Index. The leadership claim rests at least as much on the refusal behaviour and specialised tooling as on that single ranking.

Mistral also disclosed that monitoring of live traffic showed the model refusing black-box attacks on real targets, but that it sees potential for attackers to gain uplift from the model and is adjusting its guardrails accordingly. Enterprise partners including Orange Cyberdefense are receiving preview access to a specialised security variant tuned for penetration testing, vulnerability assessment and red-teaming, with access gated behind vetting and contracts that restrict testing to owned or explicitly authorised systems.

Enterprise and browsing gains

On finance workflows measured by FinWorkBench (Finch), the model climbed to 73.3%, second among open models behind Kimi K3's 77.3%. Browsing performance reached 84.5% on BrowseComp, up from a 75% supervised fine-tuning checkpoint and ahead of GLM-5.3 and DeepSeek V4.1 Flash on the same chart.

Sovereign AI positioning

Mistral stresses that the model was trained from scratch on European infrastructure, with a documented lineage from inception to release. Combined with open weights and self-hosted deployment, that is pitched at organisations that want to own, customise and control their AI — Mistral's definition of sovereign AI. Customers can fine-tune the model on proprietary security data and policies, or distil it into smaller task-specific variants, using Mistral Forge.

Why it matters

A trillion-parameter model arriving with fully open weights is a significant marker for the open-model frontier: the gap between gated proprietary systems and downloadable ones keeps narrowing on coding, the workload where developers are most willing to switch. The token-efficiency numbers matter as much as raw scores, since agent economics, not just accuracy, decide what runs in production. The security angle cuts both ways — a capable model that handles defensive tasks without excessive refusals is genuinely useful to security teams, but Mistral's own admission of uplift potential shows the dual-use risk is unresolved. Finally, a top European lab releasing frontier-scale weights under a sovereignty banner gives enterprises and governments a credible alternative to US-hosted models, even if all benchmark figures here remain vendor-reported.