Mistral Large 4 Open Weights: 1 Trillion Parameters, $0.70/$2.10 API

On 29 October 2026 Mistral AI published the weights of Mistral Large 4 on Hugging Face, three weeks after a limited preview on 6 October, Fortune reported. The model, nicknamed Le Chonk, is a 1-trillion-parameter Mixture-of-Experts with about 52 billion active parameters and a 1M-token context window. Mistral calls it the best open model built in the West. On Mistral's own API it costs $0.70 per million input tokens and $2.10 per million output tokens, according to its model page. For teams that want a frontier-class model they can host themselves without depending on a Chinese lab, this is the release to evaluate.

What happened

The model card describes a granular MoE with 128 routed experts, 4 active per token, plus one shared expert. That gives 49 billion active parameters per token, or 52 billion counting embeddings and output layers. It takes text and image input, produces text, and supports three reasoning efforts (low, medium and the default, high). Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres. Pierre Stock, Mistral's VP of science, told Fortune it ran on Mistral's own 10-megawatt cluster in France.

The open-weight release is not the same checkpoint as the preview. Stock said, per SiliconANGLE, that the open-weight version "continues reinforcement learning training and delivers substantial gains". On Terminal-Bench 4.0 that took it from 28% to 41%.

Arthur Mensch, Mistral's chief executive, framed the release as strategic: "Open source is essential to ensure state-of-the-art AI does not end up concentrated in the hands of a few actors," he said, per SiliconANGLE.

The numbers

For scale, Fortune notes that Anthropic's Mythos 5.1 scored 60.9% on the same Terminal-Bench 4.0, and Mistral's 19.7% on FrontierCode 2.0 compares with 49.1% for Claude Fable 5.2 in Anthropic's launch table. Large 4 is a strong open model, not a frontier closed model.

The license is the real story

Mistral released Large 4 under the Mistral Ecosystem Open-Weight License, described on the card as a modified MIT license for commercial and non-commercial use, with one carve-out: "operating a Model-as-a-Service business requires a separate commercial license with Mistral." In practice you can run it inside your own products and for internal use, fine-tune it and ship derivatives. You cannot stand up a public inference API for it without a deal.

Mistral is not alone in this. Moonshot's Kimi K3.1 Pro license also singles out Model-as-a-Service, though only above $20 million in trailing revenue. I wrote about this trend in Open Weights in 2026. "Open weights" increasingly means "self-host freely, resell carefully", and your legal team should read the license before your platform team writes the Helm chart.

Why it matters

Mistral is pitching continuity. Stock told Fortune that "business continuity" is the top request from regulated-industry customers, and argued that companies should question whether Chinese open models can be relied on given possible export controls. The same Fortune piece recalls that a US export-control directive in June forced Anthropic to suspend access to Fable 5 and Mythos 5 for nearly three weeks. Weights you hold cannot be switched off by a vendor or a regulator, and a European origin answers a procurement question that DeepSeek, Kimi and GLM cannot.

What sceptics say

The strongest counter-evidence is in Mistral's own card: on the Artificial Analysis Intelligence Index v5, "the leading open-weight models from China score higher". Large 4 wins on coding token efficiency, not on overall capability.

Cost per task is the second problem. In the preview, Artificial Analysis rated Large 4 roughly on par with GPT-6 Luna and DeepSeek V4.1 Flash but found it cost more than four times as much per task as open models of similar intelligence, Fortune reported. Mistral says the final checkpoint uses fewer tokens, but that needs re-measuring independently.

Third, a trillion parameters is a trillion parameters. Even with only 52 billion active, every expert must sit in GPU memory. The card's reference vLLM command uses --tensor-parallel-size 8, and the NVFP4 checkpoint targets Blackwell GPUs. This is a one-node-of-eight-big-GPUs model at minimum, not a workstation model.

What it means for engineering and DevOps teams

My view is that Large 4 is worth a serious evaluation if your constraint is jurisdiction or vendor independence, and probably not if your constraint is cost per task.

• Start on the API, then decide on self-hosting. At $0.70 and $2.10 the hosted model is cheap enough to test quality before buying GPU time. My break-even analysis in Self-Hosting LLMs vs API Costs in 2026 applies directly.
• Use the official serving path. Mistral publishes a dedicated vLLM image and parser flags. The vLLM on Kubernetes guide covers the cluster side.
• Plan storage and pull time. At roughly 1.06 TB for the full repository, cold-starting a node from Hugging Face is a capacity event. Mirror the weights to object storage close to your GPUs.
• Get the license reviewed. Especially if you run an internal platform that other business units consume as an API.

# Serving command from the Mistral Large 4 model card (vLLM, one 8-GPU node).
# The image tag is Mistral's dedicated build; pin it rather than using :latest.
docker run --runtime nvidia --gpus all \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -p 8000:8000 --ipc=host \
  vllm/vllm-openai:mistral-large-4 \
  --model mistralai/Mistral-Large-4-1T-A52B \
  --tensor-parallel-size 8 \
  --tool-call-parser mistral --reasoning-parser mistral \
  --enable-auto-tool-choice --gpu-memory-utilization 0.85

FAQ

Is Mistral Large 4 open source?

It is open-weight, not open source in the OSI sense. The Mistral Ecosystem Open-Weight License allows commercial use, modification and fine-tuning, but operating a Model-as-a-Service business needs a separate commercial license from Mistral.

What hardware do I need to run Mistral Large 4?

Mistral's reference vLLM setup uses tensor parallelism across 8 GPUs, and the NVFP4 checkpoint targets NVIDIA Blackwell. The full repository is about 1.06 TB, so plan for a multi-GPU node with very large aggregate memory.

Is Mistral Large 4 better than Chinese open models?

On the Artificial Analysis Coding Agent Index it effectively ties GLM-5.3 while using 47% fewer tokens per task, according to Mistral. On the broader Artificial Analysis Intelligence Index, Mistral's own card says the leading Chinese open-weight models score higher.

Related: Open Weights in 2026: Chinese Labs Ship the Frontier, the Licenses Grew Clauses, and Brussels Started Enforcing, LLM Inference Engines 2026: vLLM vs SGLang vs TensorRT-LLM vs Dynamo and How to Deploy vLLM on Kubernetes with GPUs: Step-by-Step (2026).