Mistral releases Large 4 weights, a 1T-parameter model pitched at cyber defenders
Mistral has made its 1-trillion-parameter Large 4 model downloadable, pitching it at coders and security teams while warning that guardrail updates won't reach copies already in the wild.
Mistral AI has released the weights of Mistral Large 4, its 1-trillion-parameter flagship model, so companies and governments can download it and run it on their own hardware. The Paris-based lab, which nicknamed the model "Le Chonk," is pitching it first at software teams and security defenders, three weeks after putting it into a guarded public preview.
The release lands as the gap between open and closed AI keeps narrowing. Many of the open-weight models competing near the frontier have come from Chinese labs, such as Moonshot, whose Kimi K3.1 Pro was heading toward an open release earlier this week. Mistral is offering European buyers a model trained from scratch in Europe that they can inspect, fine-tune and keep running even if a vendor relationship ends.
What is inside the model
According to the model card on Hugging Face, Large 4 is a mixture-of-experts model with about 1 trillion total parameters and 49 billion active per token, or 52 billion counting the embedding and output layers. Each token is routed to four of 128 experts plus one shared expert, which keeps the compute per token far below what a dense model of the same size would need. It accepts text and images, writes text, and supports a context window of 1 million tokens.
Mistral says it trained the model on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers. The same weights handle quick chat and long agent runs through three reasoning settings, low, medium and high, chosen per request. Hugging Face's launch post lists official FP8 and NVFP4 checkpoints, and the model card describes a built-in speculative-decoding head that Mistral measured at up to 1.9 times higher throughput, depending on hardware and workload. Serving it still takes serious hardware: the published vLLM and SGLang examples run on eight GPUs.
The coding and cyber numbers
Mistral's headline claim is that Large 4 is the best open-weight model for coding. On the Artificial Analysis Coding Agent Index it scores 48%, which the company calls effectively tied with China's GLM-5.3 while using 47% fewer tokens per task. On Terminal-Bench 4.0 it reached 41%, up 13 points from the preview, and on Cognition's private FrontierCode 2.0 test it posted 19.7%. On the broader Artificial Analysis Intelligence Index it scores 26, which Mistral says is the highest among Western open-weight models, while acknowledging that leading Chinese open models score higher.
Security is the second pitch. Mistral reports a 51% score on the Artificial Analysis Cyber Index, which tests whether a model can find, reproduce and patch vulnerabilities in a codebase. CSO Online noted that Mistral's own chart shows two open models, DeepSeek V4.1 Flash and MiMo-V2.6-Pro, scoring higher on that index. The company also says the model refused 96.6% of malicious cyber prompts drawn from the JailbreakBench, StrongREJECT and AgentHarm benchmarks, the highest rate it measured among open models.
Mistral's VP of science, Pierre Stock, framed the release around access. "Access that can't be cut off by definition because the model is open-weighted," he told TechCrunch, describing how a company or government could use the model to defend against attacks.
The safety trade-off
Mistral was unusually direct about the risks. In its announcement, the company said it had monitored cyber-related traffic during the preview and that, while the model refused black-box attacks against live targets, "we see a potential for uplift to conduct real attacks." It says it is adjusting its guardrails and will keep updating them. Stock told CSO Online that those updates will apply to Mistral's managed API and future releases, not to copies that organizations have already downloaded.
A more specialized variant, tuned for penetration testing and red-teaming, is going only to vetted design partners such as Orange Cyberdefense under contractual limits. Analysts quoted by CSO Online urged caution. Forrester's Charlie Dai said the best-open-model claim is hard to validate without the technical report, which Mistral says is coming soon but has not dated. Gartner's Niyati Daftary advised running the model as a controlled component with human sign-off before any fix ships, not as an autonomous agent.
The release also arrives as regulators look harder at who controls AI distribution. Hugging Face, which hosts the weights, is the subject of a Justice Department inquiry into Nvidia's planned acquisition. In Europe, CSO Online pointed out that the EU AI Act presumes systemic risk for models trained with more than 10^25 FLOPs. Stock said Mistral has signed the EU's General-Purpose AI Code of Practice but did not say whether Large 4 has been classified as a systemic-risk model.
Pricing and the business behind it
The license is a modified MIT license that allows commercial use, modification and fine-tuning. The one carve-out is that companies selling the model as a hosted service need a separate commercial license from Mistral, which protects its API business. That API now charges $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Microsoft says it is the first cloud provider offering the model, through Microsoft Foundry, and the model is also available as an Nvidia NIM container.
That compares with the closed labs, whose models stay behind APIs and whose economics are now on public display in filings such as Anthropic's IPO prospectus. Mistral's own backers have a stake in specific use cases. TechCrunch reported that chipmaking equipment giant ASML led its Series C and Samsung led its Series D in September at a €21 billion valuation, and Stock listed chip design among the model's target uses alongside cybersecurity and finance.
Mistral says Large 4 will serve as the base for a new generation of specialized and optimized models, and that the technical report will cover its cyber evaluation methods, refusal testing and preview monitoring results.
