Mistral releases the weights of Large 4, a 1-trillion-parameter open model

Mistral released the weights of Mistral Large 4 on October 29, making good on a promise it made when the model entered public preview on October 6. The checkpoint is on Hugging Face under a modified MIT license, and the version being released is not the preview: Mistral says it spent the past three weeks on additional reinforcement learning, aimed mainly at agentic coding and defensive cybersecurity.

That makes it one of the largest open-weight models anyone can download today. It also comes with a licensing condition and a set of vendor-reported benchmarks that deserve a closer look before anyone swaps it into production.

What was released

According to the model card, Mistral Large 4 is a Mixture-of-Experts model with about 1 trillion total parameters and 49 billion active per token (52 billion counting embeddings and output layers, which is where the "A52B" in the repository name comes from). Each token is routed to 4 of 128 experts plus one shared expert. It accepts text and images, outputs text, and supports a 1M-token context window. Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers.

• Weights in FP8 and NVFP4; the NVFP4 checkpoint needs NVIDIA Blackwell GPUs
• Day-0 support in Transformers, vLLM, SGLang and llama.cpp, with community GGUF quantizations, per Hugging Face's integration post
• A built-in draft head for speculative decoding that proposes three tokens at a time
• Three reasoning levels: low, medium and high (the default). Low replaces the old "none" setting, so reasoning cannot be switched off entirely
• API pricing of $0.70 per million input tokens, $0.07 for cached input and $2.10 for output, effective from the weights release

Mistral says a full technical report is "coming soon". Until it appears, details on training data and post-training methodology are limited to what the model card and launch post state.

The license has one important condition

The Mistral Ecosystem Open-Weight License grants MIT-style rights to use, modify, fine-tune and redistribute the model, including commercially. The exception: any company that operates a "Model as a Service" business, meaning it gives third parties meaningful control over inference or post-training through an API or hosted endpoint, needs a separate commercial license from Mistral. Internal use is exempt, and so are end-user products where the model is embedded behind a specific feature. In practice, a company running it on its own servers for its own staff or apps is covered; a cloud provider reselling it as an endpoint is not.

What the benchmarks say, and who ran them

Every figure below comes from Mistral's launch post. Several cite Artificial Analysis, Cognition or Surge as the evaluator, but Promptea could not open those evaluators' own pages to confirm them, and Hugging Face's integration post quotes some different numbers that appear to match the preview. Treat them as vendor-reported until independent results are published.

• Coding: 48% on the Artificial Analysis Coding Agent Index, which Mistral describes as effectively tied with GLM-5.3 while using 47% fewer tokens per task. Terminal-Bench 4.0 rose 13 points from the preview to 41%, with 30% fewer output tokens.
• Code quality: 19.7% on Cognition's private FrontierCode 2.0 benchmark, the highest among the open models Mistral charted.
• Human preference: in Surge's blind pairwise coding comparisons, Mistral says graders preferred Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, but chose Claude Opus 5.5 over it 78% of the time.
• Cybersecurity: 51% on the Artificial Analysis Cyber Index, and an average 96.6% refusal rate on malicious cyber prompts drawn from JailbreakBench, StrongREJECT and AgentHarm.
• Agent security: 97.1% attack resistance on the B3 benchmark, up from 93.3% in the preview.

Note how Mistral frames the claims. It says Large 4 is the best open-weight model for coding, not the best model overall, and its own human-preference chart shows a clear gap to the closed frontier.

The cybersecurity caveat

Mistral pitches Large 4 to defenders: vulnerability triage, reproduction and patching on infrastructure an organization controls. The launch post also includes an unusually candid admission. After monitoring cyber-related traffic during the preview, Mistral says it saw "a potential for uplift to conduct real attacks" and is changing its guardrails, with more updates planned over the coming months.

That tension is built into the release. Guardrail updates apply to Mistral's API; weights that are already downloaded run whatever safety behavior they shipped with. Mistral is separately giving vetted partners such as Orange Cyberdefense preview access to a specialized cyber variant for penetration testing and red-teaming, under contractual restrictions. That variant is not part of the open release.

What changes for developers

For teams that use coding agents, the claim worth testing is token efficiency rather than peak score. If Large 4 really reaches similar Coding Agent Index results with roughly half the tokens of its nearest open rival, the cost per completed task drops even at the same price per token. The reasoning levels are the main control: Mistral recommends low for simple tasks and keeps high as the default, so leaving it untouched means paying for maximum reasoning on every call. The model card also notes that max_completion_tokens counts reasoning tokens as well as visible output, so budgets set for non-reasoning models can cut answers short.

Self-hosting is realistic only for organizations with serious hardware. Hugging Face's ready-made deployment recipe uses eight NVIDIA H200 GPUs, and the lighter NVFP4 version requires Blackwell-generation cards. For most developers, the open weights matter less as something to run locally than as a guarantee: the model can be audited, fine-tuned and kept running regardless of what happens to Mistral's API.