Mistral Large 4 Open Weights: What It Takes to Run Le Chonk

When Mistral previewed its trillion-parameter flagship in early October, the open weights were a promise. On October 29, 2026, the promise landed. Mistral published the weights of Mistral Large 4, nicknamed "Le Chonk," on Hugging Face, and the checkpoint it released is not the preview model. Three more weeks of reinforcement learning sharpened its coding and defensive-security skills, and the API price was cut roughly in half on the same day.

• Size: about 1 trillion total parameters, with 49 billion active per token (52 billion counting embeddings and output layers), and a 1M-token context window.
• Price: $0.70 per million input tokens and $2.10 per million output tokens, down from $1.36 and $4.18 during the preview.
• Local footprint: Unsloth's smallest 1-bit GGUF is 233GB and targets a machine with 256GB of RAM; 2-bit builds run 302GB to 378GB.
• License: the Mistral Ecosystem Open-Weight license, a modified MIT license that is free for commercial use except for model-as-a-service providers.

What Changed Between the Preview and the Open-Weight Release?

We covered the original announcement in our Mistral Large 4 explainer, and the architecture is the same: a mixture-of-experts design with 128 routed experts, four of which fire per token alongside one shared expert. It is natively multimodal on the input side, reading text and images, and Mistral says it is fluent in more than 160 languages.

What is new is the training. According to Mistral, the extra reinforcement learning lifted the model's Terminal-Bench 4.0 score by 13 percentage points to 41% while cutting output tokens by 30%. It also added three selectable reasoning modes. Low is for simple tasks, Medium for everyday work, and High, the default, for the hardest problems where precision matters most. Being able to dial reasoning effort per request is a practical win, because a quick chat reply no longer has to pay for a long chain of thought.

How Does Mistral Large 4 Compare With Other Open Models?

All benchmark figures here come from Mistral's own announcement and should be read as the company's claims until independent testers reproduce them. With that caveat, the headline numbers are strong:

• Artificial Analysis Coding Agent Index: 48%, which Mistral describes as effectively tied with GLM-5.3 while using 47% fewer tokens per task.
• Cognition's FrontierCode 2.0: 19.7%, the top score among the open models Mistral listed.
• Artificial Analysis Cyber Index: 51% on finding, reproducing and patching vulnerabilities.
• BrowseComp: 84.5% on web-browsing research tasks.

The security story is the most interesting part. Mistral says the model handles legitimate defensive work without needless refusals while declining 96.6% of malicious cyber requests across standard jailbreak test suites. It also resisted 97.1% of attacks on the B3 agent security benchmark, up from 93.3% in the preview. For teams building AI agents that touch real systems, that resistance to prompt injection matters as much as raw coding ability.

Can You Run Mistral Large 4 Locally?

Yes, if "locally" means a serious workstation or server. Mistral provides the full weights in FP8 and a separate NVFP4 checkpoint aimed at NVIDIA Blackwell GPUs, with recipes for vLLM and SGLang. A multi-GPU node remains the comfortable home for this model.

For everyone else, Unsloth published dynamic GGUF quantizations within days. Its model card lists the 1-bit build at 233GB for a 256GB RAM machine, 2-bit builds from 302GB to 378GB for systems with 384GB to 512GB, and a near-full-precision Q8 at 1.1TB. Because only about 49 billion parameters are active per token, a high-memory CPU box or a unified-memory workstation can generate text at usable speeds even when the whole model never touches a GPU. Unsloth also ships a DSpark speculative-decoding draft head that it says roughly doubles decode speed at batch size 1.

If you are pricing out hardware for big local LLMs, our AI coverage tracks the memory-heavy machines that make this kind of self-hosting possible.

Why the Le Chonk Release Matters for Open-Weight AI

A trillion-parameter model with frontier-class coding, a permissive license and day-one quantized builds is a meaningful step for open-weight AI. Organizations that cannot send data to an outside API now have a top-tier option they can inspect, fine-tune and run on their own infrastructure. Mistral is also offering it through Microsoft Foundry, NVIDIA NIM containers, OpenRouter and its own API, so teams can start hosted and move in-house later. It follows Mistral's recent Voxtral 3 voice agent SDK, and together they show a company shipping open models at a steady clip.

Sources: Mistral AI — October 29, 2026; Hugging Face model card — October 29, 2026; Unsloth GGUF model card — October 2026.