Mistral AI 發佈 Mistral Large 4 權重並上架 Hugging Face

Mistral AI on October 29, 2026 released the weights of Mistral Large 4, its flagship 1-trillion-parameter natively multimodal model known as Le Chonk, making the checkpoint downloadable from Hugging Face for self-hosting and available across its API and partner platforms under a modified MIT open-weight license.

The release follows the model’s public preview on October 6, 2026, when Mistral first introduced the 1.05-trillion-parameter mixture-of-experts system. In its announcement, Mistral said it had applied feedback from the preview while continuing to scale reinforcement learning, producing what it described as major gains in coding, defensive cybersecurity, and reasoning efficiency. The company characterized Mistral Large 4 as the best open-weight model for coding and said the gains were continuing to accelerate with no signs of saturation.

License, Pricing, and Availability

The weights are published on Hugging Face under the Mistral Ecosystem Open-Weight License, a modified version of the MIT license. Under its terms, the model is free to use, modify, fine-tune, and commercialize, with one condition: businesses operating a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially.

API access is priced at $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens, according to Mistral’s model documentation. The model is served as mistral-large-4 through the Mistral API and Mistral Studio, and it is also available through OpenRouter, Mistral’s Vibe Code coding assistant, NVIDIA-hosted endpoints with an optimized NVIDIA NIM container, and Microsoft Foundry. Mistral said private deployment options are offered for Azure developers through Foundry, and Microsoft’s Tina Schuchman, corporate vice president of Foundry, said Microsoft is the first cloud provider to bring Mistral Large 4 to customers.

Coding Benchmarks and Reasoning Modes

Mistral reported that Mistral Large 4 leads 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, with a score of 19.7%, ahead of listed open-weight models including GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash.

On the Artificial Analysis Coding Agent Index, Mistral reported a score of 48%, describing the result as 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, the company reported an improvement of 13 percentage points over the public preview to 41%, taking the lead among open-weight models while reducing output tokens by 30%. In Surge Code blind pairwise human-preference evaluations cited by Mistral, 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 supports three reasoning modes: low, medium, and high, with high as the default. Mistral said the low mode replaces the previous “None” reasoning mode and is suited to simple tasks, while medium covers everyday use and high is intended for the hardest use cases requiring thorough reasoning, letting developers trade performance against cost.

Cybersecurity Results and Guardrail Changes

Mistral reported a 51% score on the Artificial Analysis Cyber Index, which measures a model’s ability to discover vulnerabilities in a codebase, reproduce and validate them, and patch them without breaking existing functionality. The company said the model handles legitimate defensive security tasks without unnecessary refusals while its built-in guardrails distinguish between defensive and malicious offensive tasks. Mistral reported an average refusal rate of 96.6% on malicious cyber requests across the cyber-related prompts of JailbreakBench, StrongREJECT, and AgentHarm, describing it as the highest it measures among open models.

The announcement also disclosed that Mistral monitored cybersecurity-related use in its live traffic following the preview. The company said it observed the model refusing to comply with black-box attacks against live targets, but that it sees potential for uplift to conduct real attacks and is modifying its guardrails accordingly, with further updates promised in the coming months. Enterprise design partners such as Orange Cyberdefense are receiving preview access to a specialized variant of Mistral Large 4 optimized for penetration testing, vulnerability assessment, and red-teaming, subject to vetting and contractual safeguards that restrict testing to owned or explicitly authorized systems in controlled environments.

The announcement included evaluation statements from executives at Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, Elastic, ENISA, Fortinet, NVIDIA, Orange Cyberdefense, Rubrik, and Zscaler. Brad Medairy, president of national cyber at Booz Allen Hamilton, said testing showed the model matching the vulnerability-detection recall rates of other frontier cyber models against the firm’s internal benchmark for source-code security analysis.

Architecture and Training

According to the model card, Mistral Large 4 is a granular mixture-of-experts model with roughly 1 trillion total parameters and about 52 billion active parameters per token (49 billion excluding embeddings and output layers), with 128 routed experts, four active per token, plus one shared expert. It supports a 1-million-token context window and accepts text and image input with text output through a 1.6-billion-parameter Pixtral-style vision encoder trained to handle variable image sizes and aspect ratios. Mistral states the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in the company’s own datacenters in Europe and is natively fluent in more than 160 languages, including every official language of the European Union.

The checkpoint ships with an embedded DSpark speculative-decoding draft head, which proposes blocks of three tokens and which Mistral measured at up to 1.9x higher throughput than serving the target model alone; a quantized NVFP4 checkpoint targeting Blackwell GPUs is also available. On the Artificial Analysis Intelligence Index v5, the model card reports a score of 26, described as the highest among western open-weight models while noting that leading Chinese open-weight models score higher on the aggregate index. Mistral also reported its Finch (FinWorkBench) score climbing to 73.3%, second among open models to Kimi K3; BrowseComp reaching 84.5% from a supervised fine-tuning checkpoint that scored 75%; and attack resistance of 97.1% on the B3 Agent Security Benchmark, up from 93.3% in the preview.

Mistral said Mistral Large 4 will serve as the foundation for a new generation of specialized and optimized Mistral models, pointing to recent releases such as Voxtral and Timestral.