Mistral Large 4 open weights: architecture, effort and license

Mistral Large 4 is a 1-trillion-parameter open-weight model from Mistral AI, with 52B parameters active per token, a 1M-token context window and text plus image input. Mistral previewed it on its API on October 6, 2026, and released the weights on Hugging Face on October 29, 2026, under a modified MIT license called the Mistral Ecosystem Open-Weight License. The hosted version costs $0.70 per million input tokens and $2.10 per million output tokens.

The released weights are not identical to the preview. Mistral says it ran about three more weeks of reinforcement learning after the preview, so the open checkpoint is the improved model. Here is how it is built, how the reasoning control works, what it takes to run, and what the license allows.

What is Mistral Large 4?

It is a general model aimed at agentic work: long tasks where the model calls tools, reads results and keeps going. The headline numbers:

Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs. It supports function calling, JSON output and strict JSON schema enforcement, and on the hosted API it also works with batching, the Conversations endpoint and built-in tools.

How does the mixture-of-experts design work?

A mixture-of-experts (MoE) model replaces each large feed-forward block with many smaller ones, called experts, and a small router that picks which experts process each token. Only the chosen experts run, so compute per token tracks the active parameters, not the total.

Large 4 uses a granular design: 128 routed experts per layer, of which the router picks 4 for each token, plus 1 shared expert that every token passes through. The shared expert holds knowledge that almost every token needs, which frees the routed experts to specialise.

The practical effect has two sides:

• Compute behaves like a ~50B dense model. Generating each token costs roughly what a model of that size would cost.
• Memory behaves like a 1T model. All 128 experts must be loaded, because the next token might need any of them.

That is why the speed is reasonable but the hardware bill is not small. More on that below.

How do reasoning effort levels work in Mistral Large 4?

Large 4 is a reasoning model: it can think before answering, and you control how much with a reasoning_effort parameter set to low, medium or high. The default is high.

Mistral's guidance maps them to work types. low is for simple tasks and fast chat replies, and it replaces the older mode with reasoning switched off. medium is for everyday use. high is for agentic loops and hard problems where precision matters. If you have used effort levels on Claude Haiku 5.5, the idea is the same: one model, a dial for how much thinking you pay for.

Two details matter when you budget tokens. First, max_completion_tokens counts reasoning tokens as well as the visible answer, so a low cap can cut off a high-effort response before it says anything. Second, Mistral recommends the same sampling settings at every effort level: temperature=1.0 and top_p=0.95.

The API is OpenAI-compatible, so the standard client works against a local server:

from openai import OpenAI

client = OpenAI(api_key="EMPTY", base_url="http://localhost:8000/v1")

response = client.chat.completions.create(
    model="mistralai/Mistral-Large-4-1T-A52B",
    messages=[{"role": "user", "content": "Summarise this changelog in five bullets."}],
    reasoning_effort="medium",
    temperature=1.0,
    top_p=0.95,
    max_completion_tokens=4096,
)
print(response.choices[0].message.content)

When served with vLLM's Mistral reasoning parser, the thinking comes back in a separate reasoning field on the message, so it does not mix with the answer you show users.

What does it take to run locally?

The weights ship in two main formats:

• FP8, the main checkpoint, an 8-bit floating-point format.
• NVFP4, a 4-bit floating-point format for NVIDIA Blackwell GPUs (B200, B300, GB200, GB300). It quantises the experts to 4 bits and keeps the other linear layers in FP8. Mistral expects it to perform close to the FP8 checkpoint.

Some quick arithmetic shows the scale. A trillion parameters at one byte each is about 1 TB of weights before any KV cache, the per-token memory a model keeps for context it has already read. The model card's examples use 8 GPUs with tensor parallelism for FP8, and 4 Blackwell GPUs for NVFP4 on SGLang.

The recommended engines are vLLM and SGLang, both with dedicated Docker images. A trimmed vLLM command:

docker run --runtime nvidia --gpus all --ipc=host -p 8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  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

Mistral also ships a small draft model for speculative decoding, a technique where a cheap model guesses the next few tokens and the large model checks them in one pass. Enabling it with three speculative tokens is reported to give up to 1.9x throughput, depending on hardware and workload. Transformers, llama.cpp (through GGUF files), Unsloth, Axolotl and TRL are supported too.

If you do not have that hardware, the hosted route is simpler. The API charges $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. The cached rate matters for agents, which resend the same long system prompt and tool list on every turn; the same logic as prompt caching on GPT-6 applies. The model is also offered through NVIDIA endpoints, Microsoft Foundry and OpenRouter.

What does the Mistral Ecosystem Open-Weight License allow?

The license is a modified MIT license. You can use, copy, modify, fine-tune, distribute and sell the model and derivatives, provided you keep the copyright notice.

There is one condition. If you or an affiliate run a Model as a Service business, meaning you give third parties inference or post-training access to the model with meaningful control over inputs, parameters or training data, for example a public API, you need a separate commercial license from Mistral for commercial use.

The license spells out what does not count. Building an end-user product where the model sits inside specific features is fine. So is relaying requests to a model someone else hosts, internal use, and anything accessed through Mistral's own products. For most teams building an app on top of the model, the MIT-style terms are what apply.

Key takeaways

• Mistral Large 4 is a ~1T-parameter MoE with 52B active, 1M context and image input, open-weight since October 29, 2026.
• reasoning_effort (low, medium, high, default high) trades depth for speed, and reasoning tokens count toward max_completion_tokens.
• Compute scales with the 52B active parameters, but memory scales with the full trillion; plan for a multi-GPU node.
• The hosted API is $0.70 in and $2.10 out per million tokens, with a 90% discount on cached input.
• The license is MIT-style except for companies offering the model itself as a hosted service.

FAQ

Is Mistral Large 4 free to use commercially?

Yes for most uses. The license allows commercial use, modification and fine-tuning. Only businesses that offer the model as a hosted service to third parties need a separate license.

Can Mistral Large 4 read images?

Yes. It accepts text and images as input through a 1.6B vision encoder that handles varied image sizes and aspect ratios. Output is text only.

Which reasoning effort should you use?

Use low for quick replies, medium for everyday tasks and high for agentic loops and hard problems. Start lower and raise it only where answers fall short, since higher effort costs more tokens.