Three Weeks of Extra RL Later, Mistral Hands Over the Weights to Its Trillion-Parameter Le Chonk

When Mistral put Mistral Large 4 into public preview on October 6, it promised open weights by the end of the month. On October 29 it delivered, and the checkpoint on Hugging Face is not the one developers tested three weeks ago. Mistral says it kept scaling reinforcement learning on preview feedback, and the released model, nicknamed Le Chonk, now carries the company’s claim of being the best open-weight model for coding.

The weights sit at mistralai/Mistral-Large-4-1T-A52B, with a separate NVFP4 checkpoint aimed at Blackwell GPUs that Mistral says has a smaller memory footprint and the same performance. The model card lists vLLM as the recommended runtime, with SGLang, Transformers and llama.cpp GGUFs also supported, and Axolotl and Unsloth guides for fine-tuning.

Inside the trillion-parameter checkpoint

According to the model card, Mistral Large 4 has about 1 trillion total parameters. It activates 49 billion per token, or 52 billion counting embeddings and output layers, through 128 routed experts with four active per token and one shared expert. It takes text and image input through a Pixtral-style vision encoder, produces text, and supports a 1-million-token context. Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres and is fluent in more than 160 languages, including every official EU language.

Reasoning effort is now set per request in three levels: low, medium and high, with high as the default. The blog says low replaces the earlier “None” mode.

One practical addition is in the checkpoint itself. It ships with an embedded DSpark speculative-decoding draft head, made of three dense MLA draft layers in BF16 that propose blocks of three tokens. Mistral says DSpark does not change the target model’s outputs and measured up to 1.9x higher throughput than serving the model alone, adding that gains depend on hardware and workload. The sample vLLM command on the card uses tensor parallelism across eight GPUs.

Coding scores that moved since the preview

All of the following figures come from Mistral. On Terminal-Bench 4.0, the company says Mistral Large 4 rose 13 percentage points from the preview to 41%, the lead among open-weight models, while producing 30% fewer output tokens. On the Artificial Analysis Coding Agent Index it reports 48%, “effectively tied” with GLM-5.3 while using 47% fewer tokens per task on average. On Cognition’s private FrontierCode 2.0 benchmark, which asks whether a maintainer would merge the agent’s change, Mistral reports a leading 19.7%.

The picture is less one-sided on general intelligence. The model card puts Mistral Large 4 at 26 on the Artificial Analysis Intelligence Index v5, which Mistral calls the highest among Western open-weight models, while acknowledging that leading Chinese open models score higher on that aggregate.

Mistral also reports 84.5% on BrowseComp, up from 75% at the supervised fine-tuning checkpoint, and 73.3% on the Finch finance-workflow benchmark. On the B3 agent security benchmark, it says the model resisted 97.1% of attacks, up from 93.3% at preview.

Defensive security, and a guardrail caveat

Mistral is pitching the model hard at cyber defenders. It reports 51% on the Artificial Analysis Cyber Index and an average 96.6% refusal rate on malicious cyber prompts drawn from JailbreakBench, StrongREJECT and AgentHarm, which it calls the highest it measures among open models.

The blog is also candid about a limit. Since the preview, Mistral says it has monitored cyber-related traffic and, while the model refused black-box attacks on live targets, “we see a potential for uplift to conduct real attacks.” It says it is modifying guardrails and will keep updating them. A specialised variant for penetration testing and red-teaming is going only to vetted enterprise design partners such as Orange Cyberdefense, under contractual safeguards.

The licence, the price and what is still missing

The weights ship under the Mistral Ecosystem Open-Weight License, which the company describes as a modified MIT licence. Commercial and non-commercial use, modification and fine-tuning are allowed. The one condition is that businesses running a model-as-a-service offering need a separate commercial licence from Mistral.

On Mistral’s API the model costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. It is also on OpenRouter, NVIDIA-hosted endpoints and NIM, and Microsoft Foundry, which Microsoft says makes it the first cloud provider to offer the model.

Mistral says a full technical report is “coming soon.” Until then, every benchmark above is the company’s own, and independent runs on the released weights are the next thing to watch.