Mistral has published the weights for Mistral Large 4, codenamed Le Chonk, moving its flagship 1-trillion-parameter model from API-only preview to self-hosted deployment. The company's release post describes it as a "natively multimodal model" trained from scratch on European infrastructure, and the weights are downloadable from Hugging Face today under what Mistral calls the Mistral Ecosystem Open-Weight license, a modified MIT license.
The architecture is a hybrid instruct-and-reasoning mixture-of-experts model that Mistral says unifies instruction-following, reasoning, and agentic capability in one checkpoint, with native multimodal input and fluency in more than 160 languages, according to Mistral's release post. Developers get three selectable reasoning modes — Low, Medium, and High (default) — letting teams trade inference cost against depth of reasoning per request rather than switching model families. Mistral recommends Low for simple tasks, Medium for everyday work, and High for cases "requiring high precision and thorough reasoning," per the post.
For self-hosting, Mistral points to vLLM and SGLang as the serving stacks, with an NVIDIA NIM container available for teams running on NVIDIA-hosted endpoints, plus the open-source NeMo AutoModel library for customization and TensorRT-LLM for high-throughput serving. Mistral also runs it as a hosted option: Microsoft Foundry became, according to the post, the first cloud offering of Mistral Large 4, with private deployment options for Azure developers, which Mistral's post frames as sovereignty without mandatory self-hosting.
On cost, Mistral's own API pricing for the model, as stated in the release, is $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. That's the number a team evaluating against closed APIs can plug directly into a cost model this week, though it describes Mistral's hosted price, not the cost of running the open weights on owned GPUs. On throughput efficiency, Mistral's post reports the model scoring 48% on the Artificial Analysis Coding Agent Index — "effectively tied" with GLM-5.3, the leading open-weight model by that measure — while using 47% fewer tokens per task on average, a token-efficiency gap that matters directly for self-hosted inference cost per request.
Benchmark gains since the preview were substantial by Mistral's account: on Cognition's private FrontierCode 2.0 benchmark the model scores 19.7%, which Mistral's post says leads open-weight models; on Terminal-Bench 4.0 it improved 13 percentage points to 41% versus the public preview while cutting output tokens by 30%. On security, Mistral reports a 51% score on the Artificial Analysis Cyber Index and a 96.6% average refusal rate on malicious cyber requests across JailbreakBench, StrongREJECT and AgentHarm prompts — the figures Mistral cites as evidence the model distinguishes defensive from offensive cyber tasks without blanket refusal.
What the release does not resolve is real-world misuse risk at the edge of that defensive/offensive line. Mistral's own post states that in monitoring live traffic since the preview, the company "observed the model refusing to comply with black-box attacks against live targets" but adds "we see a potential for uplift to conduct real attacks," and says guardrails are being modified accordingly. A specialized cybersecurity-optimized variant is being given to vetted enterprise partners like Orange Cyberdefense under contractual safeguards rather than released openly — meaning the fully open weights a team pulls down this week are not the same artifact Mistral considers safe for offensive security testing. The post also notes a full technical report "is coming soon," so independent verification of the benchmark numbers above currently rests on Mistral's own write-up rather than a published methodology paper.
For an architect comparing this against a closed API, the decision this week isn't just performance parity — it's whether the token-efficiency gain and $0.70/$2.10 hosted pricing beat your own self-hosting economics on vLLM or SGLang before the technical report and independent evals catch up to the launch claims.