Pull to refresh
Logo
What Changed
Following Why Sign Up
Mistral releases open weights for trillion-parameter Large 4 model

Mistral releases open weights for trillion-parameter Large 4 model

New Capabilities

Le Chonk is downloadable, ENISA is testing it, and Mistral has built Studio agents around it while its benchmark report remains unpublished

6 days ago: Mistral Studio adds managed agents and workflows

Overview

Updated 3 days ago

Mistral's trillion-parameter flagship, nicknamed Le Chonk, can now be downloaded and run on a customer's own machines; the Paris lab published the open weights on October 29 after three more weeks of training. On November 5 it added tools to Mistral Studio for building multi-step AI agents, automated workflows and hosted apps, all built to run on the model.

The European Union's cybersecurity agency, ENISA, is testing Large 4 to find vulnerabilities, audit code and analyze defensive logs. Mistral also gave design partners including Orange Cyberdefense preview access to a security-tuned variant. Downloaded copies will not automatically receive later safety updates, and the technical report behind the benchmark claims is still unpublished.

Why it matters

Businesses can run a near-frontier model on their own machines, but downloaded copies will not get later safety patches.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

~1 trillion
Total parameters (Mixture-of-Experts)
Sparsely activated: about 49 billion parameters active per token, with Mistral citing 52 billion active per task.
48%
Score on Artificial Analysis Coding Agent Index
Effectively tied with leading open model GLM-5.3 while using 47% fewer tokens per task.
63%
Cost reduction per Intelligence Index task vs October preview
Fell from $5.54 to $2.06 per task after the reinforcement learning push.
96.6%
Refusal rate on malicious cyber requests in guardrail tests
Highest measured among open models, though limited to Mistral's own test scenario.
1 million
Context window length (tokens)
Takes text and image input; generates text output.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

Log in Sign Up

People Involved

PS
Pierre Stock
Vice President of Science, Mistral
Overseeing Large 4 safety work and guardrail updates

Organizations Involved

Mistral
Mistral
AI research lab and model provider
Released open weights, rolled out Studio tools built around Le Chonk, and runs a vetted security-variant preview

Paris-based AI lab that builds large language and multimodal models.
Hugging Face
Hugging Face
AI model hosting platform
Hosts the Large 4 open weights

The open model repository where Mistral published the Large 4 weights.

Timeline

October 2026 November 2026

5 events Latest: 6 days ago
Tap a bar to jump to that date

  1. Mistral Studio adds managed agents and workflows

    Latest Product

    Mistral adds managed agents, workflows and app hosting to Mistral Studio, all built to run on Large 4. Customers can deploy in the cloud, inside their own VPC or on-premises.

  2. Large 4 writes a quantum chemistry solver from a prompt

    Demo

    Mistral says the model produced a self-consistent field solver, checked its output against existing tools, and built a draggable 3D view of a molecule forming a bond.

  3. Open weights released on Hugging Face

    Release

    Mistral publishes Large 4 weights after three more weeks of reinforcement learning, with lower per-task cost and higher coding and cyber-defense scores.

  4. ENISA says it is testing Large 4

    Response

    The EU cybersecurity agency says it is testing the model to identify vulnerabilities, audit code and analyze defensive logs. Mistral also opens a vetted preview of a security-tuned variant to design partners including Orange Cyberdefense.

  5. Large 4 enters public preview

    Launch

    Mistral launches its trillion-parameter model via API, announcing open weights planned for later in October.

Scenarios

1

Le Chonk becomes the default self-hosted enterprise agent

Likely Resolves by Q2 2027

Discussed by: VentureBeat, The New Stack

Businesses deploy Large 4 on their own infrastructure with vLLM or SGLang, treating the $2.06 per-task cost and in-house data as decisive. Microsoft Foundry and Nvidia endpoints lower the barrier. Sustained benchmark performance would lock in the position.

2

Regulators move to restrict open-weight cyber-capable models

Possible Resolves by Q2 2027

Discussed by: CSO Online, Gartner analyst Niyati Daftary

A documented misuse or high-profile incident draws regulator attention to a downloadable model with cyber capabilities. The 96.6% refusal rate only covers Mistral's test set, Daftary notes, and downloaded checkpoints will not automatically receive future patches. The European Commission or a national government enacts rules restricting such releases.

3

DeepSeek or GLM's next release dethrones Large 4

Likely Resolves by Q2 2027

Discussed by: Artificial Analysis (via VentureBeat)

Large 4 trails DeepSeek V4.1 Flash and GLM-5.3 on Artificial Analysis's Intelligence Index. A successor from either Chinese lab would likely widen the gap, leaving Mistral to compete on cost per task and Western deployment instead of raw capability.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

July 2023

Meta releases Llama 2 (July 2023)

Meta made its 70-billion-parameter language model freely downloadable, then added conditions: companies with more than 700 million monthly users needed a separate license. It became the reference for open weights with strings attached.

Then

Sparked an ecosystem of fine-tunes and put open-weight models in enterprise reach.

Now

Established the license pattern Mistral now uses for Large 4.

Why this matters now

Mistral's custom license, free except for a paid Model-as-a-Service tier, echoes Llama's approach of open weights with commercial conditions.

January 2025

DeepSeek proves open weights can undercut closed labs (January 2025)

The Chinese lab released R1, an open-weight reasoning model that matched top Western models on math and coding at a small fraction of the price. The release triggered a sharp repricing of AI inference.

Then

Put pressure on closed-model pricing across the industry.

Now

Showed raw performance and per-task cost are separate, contestable axes.

Why this matters now

Frames Large 4's position: it trails Chinese open leaders on raw scores but competes on cost per task and Western deployment.

August 2022

Stable Diffusion opens image generation (August 2022)

Stability AI released a text-to-image model anyone could download, moving AI from the API to the desktop. The release ignited the misuse-versus-openness debate that still defines open weights.

Then

Broad creative use alongside deepfake misuse; platform moderation applied only to hosted copies.

Now

Established the pattern where guardrails ship in weights but later patches reach only managed services.

Why this matters now

The same dynamic now hits cybersecurity: Large 4's guardrails are built in, but downloaded copies won't get future updates.

Sources

(12)
+6
Mistral TechCrunch Hugging Face The Verge VentureBeat The New Stack CSO Online Podimo The Next Web GitHub Crypto Briefing Pandromeda