← Back to feed

Mistral released the weights of Mistral Large 4 — internally nicknamed Le Chonk — on October 29, delivering a model that improved materially from its earlier preview. Three weeks of post-preview reinforcement learning pushed its benchmark scores across the board before the weights were made public.

What Changed Since Preview

The preview launched October 6 with competitive but not leading benchmark numbers. The OSS release is a measurably different model:

• Terminal-Bench 4.0: 41%, up 13 percentage points from preview. Best open-weight result on this benchmark.
• FrontierCode 2.0: 19.7%. Cognition’s private repository-merge benchmark, where Mistral Large 4 now leads all open-weight models.
• BrowseComp: 84.5%, up from 75% at the SFT checkpoint. Best open-weight result, above GLM-5.3 and DeepSeek V4.1 Flash 1017.
• B3 Agent Security: 97.1% attack resistance, up from 93.3% in preview. First-place among open-source models.
• AA Coding Agent Index: 48%, effectively tied with GLM-5.3. Mistral’s version uses 47% fewer tokens per task.
• FinWorkBench (Finch): 73.3%, second among open models behind Kimi K3.
• AA Cyber Index: 51%, measuring ability to discover, reproduce, and patch vulnerabilities.

On Surge Code’s blind pairwise evaluations, Mistral Large 4 was preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash on code tasks, ranking behind only Claude Opus 5.5 in the set evaluated.

Architecture

1.05 trillion parameters total, 49 billion active per forward pass. MoE architecture with sparse attention. Pixtral vision encoder for native multimodal input. 1 million token context. Fluent in 160-plus languages. Available in FP8 and NVFP4 variants on Hugging Face; GGUF conversions from Unsloth and llama.cpp.

Three reasoning modes replace the prior binary setting:

Mode	Use
Low	Simple tasks, replaces None
Medium	Everyday tasks
High (default)	Hardest tasks requiring precision

The reasoning modes affect output token count significantly. Mistral reports that the High mode at Terminal-Bench 4.0 uses 30% fewer output tokens than its preview equivalent, despite the score gain.

Cybersecurity Focus

Mistral positioned the model explicitly for defensive security. The AA Cyber Index score of 51% measures vulnerability discovery, reproduction, and patching in a controlled environment. On JailbreakBench, StrongREJECT, and AgentHarm combined, Mistral Large 4 refused 96.6% of malicious cyber prompts — the highest measured refusal rate across open models. Simultaneously, it handles legitimate defensive tasks without unnecessary refusals.

Mistral noted it observed potential for real-attack uplift in live traffic monitoring and said it is modifying guardrails accordingly. A specialized variant for penetration testing and red-teaming is in preview for enterprise partners with contractual restrictions on use.

License and Pricing

The model ships under the Mistral Ecosystem Open-Weight License, a modified MIT license. Free to use, fine-tune, and redistribute. One restriction: businesses running a model-as-a-service product must obtain a separate commercial license from Mistral before offering it as a hosted service. Non-SaaS deployment — on-premise, internal enterprise use, fine-tuned versions — has no additional requirements.

API pricing: $0.70 per million input tokens, $0.07 per million cached input tokens, $2.10 per million output tokens. Available through Mistral API, OpenRouter, Microsoft Foundry (including private deployment on Azure), NVIDIA-hosted endpoints, and NVIDIA NIM containers.

Context

Mistral Large 4 now runs as the OSS frontier competitor in the same tier as GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash — the cluster of open-weight models operating below Claude Opus 5.5 and GPT-6 Astra on aggregate benchmarks. Its Coding Agent Index score of 48% with 47% fewer tokens is the key commercial differentiator: coding agents pay per token, and the same task quality at half the token cost meaningfully changes the unit economics at scale.

A full technical report is forthcoming.