Mistral Large 4 ("Le Chonk") in Vibe: How It Compares to Claude Code

TL;DR: Mistral Large 4, which Mistral itself nicknames "Le Chonk", is the French lab's largest model: 1 trillion parameters, 52 billion active. It has been in public preview since October 6, 2026, and the weights are due by the end of the month. It is already selectable in the Mistral Vibe CLI, and we are running it right now side by side with Claude Code on the same tasks. Results will land in an update to this article.

We write code with Claude Code every day. When "Mistral Large 4 (Preview)" showed up in our Vibe model picker, the question asked itself: can an open-weight model hold up on our real work? And if it can, what does that change for a company that needs to keep its data in-house?

---

What is Mistral Large 4 ("Le Chonk")?

Mistral announced it on October 6, 2026. Here is what the official announcement says, and only that:

Three things the announcement leaves open. The weights' license is not stated. No inference hardware requirements are given. And Mistral writes that the reinforcement learning run behind this preview "is still in flight": the model you test today is not the one that will ship.

We don't repeat the benchmark scores here: Mistral published all of them, and no independent evaluation existed when we wrote this. Prices reported in the press also differ between sources. Check Mistral Studio's own pricing before you budget anything.

Is Mistral Large 4 available in Vibe?

In the CLI, yes. This is our terminal on October 22, 2026, running Vibe 2.26.1:

A few pointers, because "Vibe" now means two things:

• Vibe is the former Le Chat. Mistral renamed its assistant on May 28, 2026, with a Work mode and a Code mode.
• The Vibe CLI is the terminal coding agent. It is open source (Apache 2.0), installs with uv tool install mistral-vibe, supports MCP servers, skills and subagents, and switches models with /model.
• The model is pinned per session. The CLI's docs say the first message pins the model to the session: to try Large 4, start a new session after picking it.
• Our version runs a new engine. Vibe 2.26.1 announced a "unified harness", with --legacy-harness to fall back if something breaks.

Mistral's announcement doesn't mention Vibe. We have only checked the CLI, not the chat app.

Mistral Vibe vs Claude Code: what already differs on paper

Before any results, the two tools make different bets. On features they look a lot alike. On where the model runs, they are opposites.

For most companies the last row matters more than any benchmark score. If your data can go through an enterprise tier under contract, the choice comes down to the quality of the work. If it cannot leave your walls, only an open-weight model is on the table, and Large 4 enters that list at the top end.

Our head-to-head, in progress

We are running both agents at the same time, on the same tasks, with the same instructions. The tasks come from our real work (Go services, this Gatsby site, Terraform infrastructure), not demo exercises.

What we look at:

• does the task finish without a human fix, and do the tests pass;
• what had to be corrected when it doesn't;
• time spent and cost per task.

Update coming. We will publish the results here, with the tasks and how each one went. Since Large 4's training isn't finished, we will run a second pass once the weights are out: a result on the preview won't tell the whole story of the final model.

Our last model comparison showed that a smaller model often matches a frontier one on a specific task: 13 models across 470 tests and 4 languages. Same approach here: real tasks rather than a single score. If you use Claude for code review, see also what Claude Code Review does and costs.

Self-hosting Mistral Large 4: what it really takes

Open weights are the real news for a company. They don't make self-hosting easy. Here is the math, with a caveat: these are our orders of magnitude, not a Mistral spec sheet, since Mistral hasn't published one.

• Memory follows total parameters, not active ones. The 52 billion active parameters make each token fast to compute, but all 1 trillion have to sit in GPU memory.
• At 8-bit, the weights alone are about 1 TB. Add the memory for in-flight conversations on top. An eight-GPU server with 80 GB cards (640 GB total) is not enough. You need eight latest-generation GPUs, or a 4-bit quantized version (about 525 GB) with a quality loss you would have to measure.
• Hardware is not the whole cost. A model on your servers needs upgrades, monitoring, and someone to look after it.

That is what our sovereign AI page says: self-hosting is justified when the data demands it (health, trade secrets, confidentiality commitments to your own clients), not as a principle. Otherwise an enterprise tier in the right region covers the need for far less. Mistral's own European deployment is a middle option worth a look. On personal data, see also AI and GDPR.

What we do, and what we don't do yet

What we do. We build software with coding agents every day, Claude Code first, and we are now adding Vibe on the same repositories. For clients, we work out where each AI use case should run (an enterprise tier, a regional deployment, or a model on their own servers), then deploy and integrate what was chosen. Mistral's speech models get the same treatment: see how Voxtral 3 did on real Gulf voice orders.

What we don't do yet. We have not run Mistral Large 4 on our own GPUs: the weights aren't out. And we don't have numbers to publish on the Claude Code comparison yet. They will appear here, and not before they are measured.

Sources

• Mistral AI, Introducing Mistral Large 4, October 6, 2026
• Mistral AI, Mistral Vibe CLI repository (license, install, MCP, skills, subagents)
• L'Éclaireur Fnac, Pourquoi Le Chat devient-il Vibe ? (French)
• Siècle Digital, Mistral revient dans la course à l'IA (French), October 7, 2026