Mistral Large 4 ships 1T-parameter open weights and leads open models on Terminal-Bench

The model, nicknamed Le Chonk, entered public preview earlier in October. Mistral says three further weeks of reinforcement learning since then moved it to the front of open-weight coding, and that continued RL scaling also improved defensive cybersecurity. Reuters headlined the release as Mistral targeting cybersecurity users, and the announcement supports that reading. It carries endorsements from Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, Elastic, Fortinet, Orange Cyberdefense, Rubrik and Zscaler, along with ENISA and NVIDIA. Mistral also says it trained the model from scratch on European infrastructure, and it pitches the model as a foundation for sovereign AI.

The retrieved text of Mistral's post calls the release fully open-weight but does not name the license, so the detail comes from the Hacker News thread. One commenter described a modified MIT license that is not OSI-approved, with the common carve-out that a business serving the model needs its own license. That commenter called the release weights-available rather than open weight, and other commenters disputed that terminology. A second commenter narrowed the restriction to a Model as a Service business, and a third quoted the exemption clause from the license text:

That commenter said they can host and run the model, would need a license to sell access as an inference provider, and compared the arrangement to the AGPL. On that reading, a team running the weights behind its own internal agents or tools falls under the exemption, and an inference provider reselling API access does not. Other commenters argued over whether terms like these are enforceable on model weights at all, but that is opinion and not settled. The excerpt does not show what Section 2 requires, or how it treats a product that exposes the model to external customers through a feature rather than a raw API. Anyone building a customer-facing product on the weights should read the full license file on Hugging Face rather than rely on the thread's summary.

The retrieved announcement text says nothing about hardware requirements, memory footprint, GPU counts or serving throughput. One Hacker News commenter noted that the download links were buried at the end of the post. They point to two Hugging Face repositories, Mistral-Large-4-1T-A52B and Mistral-Large-4-1T-A52B-NVFP4. The repository name suggests 1T total parameters with 52B active per token, but the retrieved post text does not state an active-parameter count or describe the architecture, and the NVFP4 repository name indicates a lower-precision build with no accuracy or memory figures given. NVIDIA's Kari Briski refers to a collaboration on NVIDIA accelerated infrastructure, but the quote gives no deployment numbers.

Two details in the post bear on serving cost. The model now has three reasoning modes. Low is for simple tasks and replaces the earlier None mode, Medium is for everyday tasks, and High is the default for the hardest cases requiring thorough reasoning. Mistral also says customers are tailoring the model with proprietary security data through Mistral Forge, and that for well-defined cybersecurity tasks it can be further optimized into smaller, task-specific variants, though the post does not say how small those variants are or whether they can be self-hosted. On latency, one commenter using the model for a personal project called it very slow when thinking, and the sources contain no latency numbers to check that against.

The token-efficiency figures bear directly on cost. If an agent reaches the same Coding Agent Index score with 47% fewer tokens on average, it costs roughly half as much per task at equal per-token prices. These are Mistral-reported results, and commenters flagged gaps in the comparison set. One noted that Kimi 3.1 is absent from the benchmarks. Another said they were unsure how Mistral compares with other open-weight models, particularly Chinese ones, and suggested EU provenance is the main reason to choose it.

Mistral reports 51% on the Artificial Analysis Cyber Index, which measures whether a model can find vulnerabilities in a codebase, reproduce and validate them, and patch them without breaking existing functionality. Mistral says the model handles this defensive work without unnecessary refusals. It also reports an average refusal rate of 96.6% on malicious cyber requests, measured on the cyber-related prompts of JailbreakBench, StrongREJECT and AgentHarm, and calls that the highest it measures among open models. The retrieved text gives no Cyber Index scores for other open-weight models, so the claim that it leads open-weight defensive security benchmarks rests on Mistral's summary. Booz Allen says the model matched other frontier cyber models on vulnerability detection recall in its internal source-code security benchmark with what it calls a compelling cost curve, but its quote gives no recall figures.

Mistral also makes a concession. It says that in live traffic since the preview it observed the model refusing black-box attacks against live targets, but still sees "a potential for uplift to conduct real attacks." It says it is modifying its guardrails and will continue updating them in the months to come, without saying whether those changes reach the released weights or only Mistral's hosted service. That matters for self-hosters, because a downloaded checkpoint is a fixed snapshot. Separately, a specialized variant for penetration testing, vulnerability assessment and red-teaming is in preview only for vetted enterprise design partners such as Orange Cyberdefense, under contractual safeguards and restricted to owned or explicitly authorized systems, and is not part of the public release.

The retrieved announcement text does not mention pricing. One Hacker News commenter, citing Mistral's pricing docs, reported $0.70 per million input tokens and $2.10 per million output tokens, against $1.36 and $4.18 for the preview, and said the drop warrants re-running evaluations that had found the preview uncompetitive. Another disputed the comparison, saying the preview price had already been cut to about $0.68 and $2.09, and called the model still too expensive and slow next to Opus 5.5. A third called the pricing very aggressive for a model this size and speculated Mistral may be serving at a loss, while a fourth noted that cached-input pricing was not stated, which decides the bill for agentic coding sessions with cache rates they put above 97%. Check Mistral's pricing page before building cost models on these figures.