Mistral Releases Large 4 Weights: Open 1T-Parameter Model for Coding

Mistral AI has published the weights of Mistral Large 4, its new flagship model. It is a multimodal system with 1T parameters: it works not only with text but with other types of data too. "Weights" are the numbers that store a trained neural network. Because they are open, you can download the model, run it on your own hardware and fine-tune it for your needs. The model was trained from scratch on European infrastructure. After the public preview, Mistral fine-tuned it with reinforcement learning (RL, where the model receives a "score" for its result and adjusts to it) on two classes of tasks: programming and cybersecurity.

What independent measurements show

The figures come from rankings by Artificial Analysis. The company independently runs models through standard test suites and publishes the comparison.

• Coding Agent Index: 48%. The index tests coding agents: the model receives a task in a real codebase, reads files on its own, makes edits and runs checks. 48% is Mistral Large 4's overall score.
• 47% fewer tokens than GLM-5.3. A token is the unit of text that the user pays for. To solve the same task, Mistral Large 4 uses almost half as many tokens as the open model GLM-5.3. This matters for agents: they take dozens of steps, and every step costs money.
• Artificial Analysis Cyber: 51%. This test evaluates three skills: finding a vulnerability in code, reproducing it (showing that the hole actually works) and writing a patch that closes it.

Separately, Mistral says the model audits code without false positives from its safety filters. The problem is a familiar one: many models refuse to analyze vulnerable code because they mistake a defender's request for an attack attempt. For a security specialist, such a refusal is lost time. Independent testing will show how well this works in practice.

Reasoning modes and pricing

To help control costs, the model has three built-in reasoning modes:

• Low: minimal reasoning, fast and cheap answers;
• Medium: a compromise;
• High: the default mode, with maximum reasoning on complex tasks.

Developers decide how many tokens to spend on "thinking." Low is enough for a one-line edit; High is better for hunting vulnerabilities in a large project.

Prices for the official Mistral API:

• $0.70 per million input tokens;
• $0.07 per million input tokens on a cache hit, ten times cheaper;
• $2.10 per million output tokens.

The cache discount benefits agents in particular. In a long session, they send the model the same context again and again: the project description, files already read, the history of steps.

Where to try it and what the license says

You don't have to run the model locally. It is available through:

• Mistral API;
• OpenRouter, a single gateway to many models;
• Vibe Code, Mistral's development environment;
• optimized NVIDIA NIM containers, ready for deployment on GPUs.

The weights are on Hugging Face in the Mistral-Large-4-1T-A52B repository. Judging by the name, the model uses a mixture-of-experts design: of the 1T parameters, about 52B are active for each token. This makes it faster and cheaper to run than a dense model of the same size. But you still have to store all the weights in full. In an 8-bit format that comes to about a terabyte, so you can't run the model without a cluster of several GPU servers.

The license is a modified MIT. Standard MIT restricts almost nothing, but Mistral has added its own terms. You can see exactly which ones in the LICENSE file in the repository. Company lawyers should read it before deployment, especially for commercial products.

What this means for engineers and businesses

Teams that need a coding agent and want to control their data now have another serious open-weights option. They can keep the model inside their own perimeter and fine-tune it on their own code. Its European origin and training on European infrastructure give companies with data-residency requirements an argument in its favor.

Evaluate the model on your own tasks. The Artificial Analysis index is a good guide, but it is no substitute for a trial run on your repository. The 47% token saving relative to GLM-5.3 delivers real benefit only at your volumes and in your scenarios. And a 51% score in cybersecurity means the model solves roughly half of the test's tasks. It does not yet replace a human security specialist, but it is already useful as an assistant for code audits.