Mistral Large 4 (Le Chonk)

What Mistral Large 4 (Le Chonk) is

Mistral Large 4, also called Le Chonk, reads text and images and generates text. It is a model to connect to an application, rather than a complete autonomous assistant. Mistral has released improved weights after an earlier public preview; its API documentation still labels the model Public Preview.

The model uses a mixture-of-experts architecture: a smaller set of parameters works on each token, while the full collection of weights still has to be stored. That distinction explains why a model described with roughly 52 billion active parameters can require far more memory than a dense model of that size.

How it fits into a workflow

For an initial trial, send a small, rights-cleared document or coding task through the API and check the answer against the original material. This is an editorial trial suggestion, not a completed test. A self-hosted application instead loads a checkpoint in a compatible runtime and supplies its own tools, permissions and review steps. Tool-call support does not automatically give the model a browser or working access to other services.

Examples described as illustrative or proposed trials are editorial suggestions. Documented real builds, when available, are linked separately below.

Requirements, cost and limitations

Self-hosting needs much more than a typical laptop GPU. The published vLLM configurations use datacenter accelerators and a model-specific runtime image. Weights, conversation state, runtime buffers and simultaneous requests all consume memory. Before commercial deployment, read the custom license: businesses operating Model-as-a-Service, including through an affiliate, face a separate-license condition; the text defines exceptions for internal use and access through Mistral products.

When to choose it and what to compare

Use API access to evaluate the workflow before committing to infrastructure. Consider self-hosting when deployment control justifies operating servers and checking the license. Compare that decision with smaller open models, not only with leaderboard positions. The new release's benchmark claims indicate tasks worth testing, but do not establish your application's reliability or a universal winner.

Access, running cost and practical requirements

Frontier server-scale candidate — editorial selection, no verified market lead. How we use this label.

Use the paid Mistral API, still labeled Public Preview in its operational docs, or download the exact FP8/NVFP4 checkpoint for compatible server infrastructure.

Inputs and outputs

Text and images in; text out.

License / access terms

Mistral Ecosystem Open-Weight License; modified MIT with commercial conditions

Context

Checkpoint advertises 1M positions; the current vLLM recipe validates up to 256K. Capacity arithmetic is not a demonstrated 1M-context serving result.

Hosted API option

Model ID

mistral-large-4

Status

Public Preview in current model documentation

Input / million tokens

$0.7

Cached input / million tokens

$0.07

Output / million tokens

$2.1

Usage billing through Mistral API. These are current listed rates, not a subscription allowance, tax-inclusive quote, or all-in workflow cost. Self-managed infrastructure has separate capacity and operating costs.

Memory and hardware planning

vLLM lists roughly 1.06 TB FP8 or 610 GB NVFP4 weights. Its validated examples include four GB300 GPUs for NVFP4, eight B200/B300 GPUs, and sixteen H200 GPUs for FP8. Weight estimates exclude context/runtime headroom; these are specific configurations, not interchangeable GPU-count minimums.

A practical setup path

As checked, vLLM prerequisites specify its model-specific image vllm/vllm-openai:mistral-large-4 and say standard release/nightly wheels do not yet contain support. Follow the recipe's current hardware-specific configuration; no installation or serving was tested here.

Consider API evaluation first; self-hosting requires substantial infrastructure and a review of the custom license.

Running downloaded weights on your own computer has no hosted-model token bill, but hardware, power and maintenance still cost money. Rented GPUs and managed APIs charge differently. Check the model license and the runtime before spending on a deployment.

Benchmark evidence and what it means

These are selected source-reported observations, not a league table. Task, version, tool access, reasoning budget and test harness all affect a result. A historical score does not establish current leadership or accuracy on your own work.

No normalized quality score is proposed. Release claims lack sufficient public harness, model revision and inference-budget detail for a controlled comparison; runtime integration tests answer a different question. Keep unknown settings blank rather than borrowing generic recommended sampling values.

Sources and evidence

Product facts are based on the following published sources, checked 2026-10-29. Workflow suggestions and comparisons are editorial interpretation; PrisAIm has not independently benchmarked this product.