Mistral releases 1-trillion-parameter Large 4 weights, halves API prices

Summary

• Mistral published downloadable weights for its trillion-parameter flagship Large 4 on October 29, 2026.
• The open-weight license permits commercial use and fine-tuning, but resale as a hosted service needs separate terms.
• Enterprises can now run a frontier-class model in-house at half the prior interface price.

The latest

Mistral made the weights of its flagship Mistral Large 4 model freely downloadable on October 29, 2026, following a three-week public preview. The hybrid mixture-of-experts system carries 1.05 trillion total parameters with 52 billion active per pass, and ships under a license the company describes as a modified MIT. Interface prices were cut by half alongside the release.

Details

• The architecture: According to Techstrong.ai, Large 4 is a hybrid instruct-and-reasoning mixture-of-experts model with a 1-million-token context window, natively multimodal across text and image inputs, and competent in more than 160 languages. It offers three reasoning levels: low, which replaces the former no-reasoning mode, medium, and high as the default.
• The training run: Techstrong.ai reported the model was trained on roughly 3,800 NVIDIA Grace Blackwell processors and previewed for three weeks with corporate and government bodies before the weights went public. The repository is listed on Hugging Face as mistralai/Mistral-Large-4-1T-A52B, nicknamed Le Chonk.
• The license terms: The Mistral Ecosystem Open-Weight license permits using, modifying, fine-tuning and commercializing the model. Companies offering it to third parties as Model-as-a-Service, or selling post-training capabilities, require a separate commercial license. Internal and embedded enterprise deployment is exempt.
• The coding scores: Mistral reported 19.7% on FrontierCode 2.0, a private Cognition benchmark, presented as the highest result shown. On the Artificial Analysis coding-agent index it scored 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task.
• Agentic gains: Mistral put Terminal-Bench 4.0 at 41%, up 13 points from the preview's 28%, with 30% fewer output tokens. BrowseComp reached 84.5% against 75% for an earlier supervised fine-tuning checkpoint, and FinWorkBench hit 73.3%, second among open models behind Kimi K3.
• Security numbers: Mistral reported 51% on the Artificial Analysis cyber index and a 96.6% average refusal rate for malicious cyber requests across JailbreakBench, StrongREJECT and AgentHarm. Resistance on the B3 Agent Security Benchmark was listed at 97.1%, up from 93.3% at preview stage.
• Human preference: In Surge Code's blind human-preference evaluation, Mistral said Large 4 placed ahead of GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Among the models assessed, only Claude Opus 5.5 ranked above it.
• The pricing: Per million tokens, Mistral listed 0.70 dollars for input, 0.07 dollars for cached input and 2.10 dollars for output. Techstrong.ai reported interface prices were cut to half their previous level alongside the open-weight release.
• Where it runs: Weights sit on Hugging Face, with hosting through Mistral Studio, Mistral's own interface, OpenRouter, NVIDIA endpoints and Microsoft Foundry, which Mistral called the first cloud provider to offer it. Self-hosting runs via vLLM and SGLang, with NVIDIA NIM, NeMo AutoModel and TensorRT-LLM support.
• The partners: Named deployment partners include Microsoft, NVIDIA, Elastic, Hugging Face and ZML. Kerry Briski, NVIDIA's vice president of open models and software, said open models let security researchers and developers inspect, adapt and build on artificial intelligence. Mistral also cites Cisco, Cloudflare, CrowdStrike, Fortinet, ENISA, ASML, Crédit Agricole and ESA.

Between the lines

The license carve-out defines the commercial strategy: by exempting internal and embedded enterprise deployment while requiring separate terms for Model-as-a-Service resale, Mistral invites banks and government bodies to run the weights themselves, while keeping rival hosting providers inside a paid channel. The token-efficiency claims — 47% fewer tokens on coding agents, 30% fewer outputs on Terminal-Bench — point at operating cost as the pitch.

What's next

Watch for independent benchmark reproductions outside Mistral's own reported figures, particularly on FrontierCode 2.0 and Terminal-Bench 4.0, and for which additional cloud providers follow Microsoft Foundry in hosting the model.

Source: Mistral, Techstrong.ai