Mistral Open-Sources “Le Chonk”: Trillion-Parameter Large 4 Weights Released, Chasing GLM-5.3

What Happened: The Weights Dropped

On October 29, 2026, French AI lab Mistral released the weights for Mistral Large 4 (ML4) — the trillion-parameter model it previewed on October 6 and unofficially nicknamed “Le Chonk.” The weights are live on Hugging Face under the model ID mistralai/Mistral-Large-4-1T-A52B, and the model is also available through Mistral’s API and Mistral Studio console.

Why does this matter? Because until today, if you wanted a top-tier open-weight model, the strongest options were overwhelmingly Chinese — DeepSeek, Qwen, GLM, and Kimi have dominated the open leaderboards for the past year. ML4 is Mistral’s bid to become the first western lab since the Llama era to genuinely compete at that level, and it arrives at a moment when enterprises and governments are actively looking for models they can run on their own infrastructure without worrying about a vendor relationship going sour.

Mistral’s framing is blunt: access that can’t be cut off, because the weights are yours. For regulated industries, defense, and public-sector buyers, that argument is landing harder every quarter.

Inside Le Chonk: 1T Parameters, 52B Active

ML4 is a natively multimodal Mixture-of-Experts model: roughly 1.05 trillion total parameters, but only about 52 billion active per token, plus a 1.6B vision encoder. That MoE design is why a model this large can be served at all — and why Mistral can price API access at $0.70 per million input tokens and $2.10 per million output tokens, with cached input at $0.07.

• 1M-token context window — long enough for entire codebases and document sets.
• Multimodal input — images, charts, technical drawings, and documents, with state-of-the-art visual grounding among open weights.
• 160+ languages, including every official EU language.
• Agentic by default — native function calling, JSON output, and strict JSON schema enforcement.
• Hybrid reasoning — instruction, reasoning, and agentic behavior unified in one model, with configurable reasoning effort (low / medium / high).

Mistral also says ML4 was trained with the same pipeline it offers customers through Mistral Forge, and that it will serve as the foundation for a new generation of specialized models in the months ahead.

The Numbers: Tied With GLM-5.3 at 47% Fewer Tokens

The headline comparison is against GLM-5.3, the open-weight model widely considered the strongest in the world before today. On independent benchmarks tracked by Artificial Analysis:

Two things stand out. First, the token efficiency claim: matching GLM-5.3’s agentic-coding score while using roughly half the tokens per task is a real economic advantage — at agentic workloads, token spend is the budget. Second, the honesty gap: ML4 tops the Intelligence Index v5 at 26 among western open-weight models, but that’s still well behind the GLM-5.3 class (and far behind closed frontier models like Gemini 4 Carbon at 59). Mistral is winning on efficiency and sovereignty, not raw intelligence.

Mistral also claims state-of-the-art open-weight performance in cybersecurity, finance, and manufacturing — and in visual grounding, where it says it surpasses even some closed frontier models.

The Cybersecurity Pitch

The timing of the release is deliberate. Google shipped Gemini 4 Carbon this week as a closed, gated frontier model aimed squarely at cyber-defense work, and Mistral is positioning ML4 as the open counterweight — a model that security teams and national agencies can deploy on-premises for vulnerability research and defense.

That matters because cyber work touches exactly the data organizations are least willing to send to a third-party API. Mistral emphasizes that ML4 is state of the art among open weights for security work while refusing malicious requests — a balancing act every security-focused model vendor now has to demonstrate. For teams comparing options, the open DeepSeek family and closed assistants like ChatGPT and Claude remain the practical alternatives, each with different privacy trade-offs.

The License Fine Print

ML4 ships under the Mistral Ecosystem Open-Weight License — a modified MIT license covering both commercial and non-commercial use, with one significant carve-out: operating a Model-as-a-Service business on top of the weights requires a separate commercial license with Mistral.

In practice: deploying ML4 inside your company, product, or government agency is fine; reselling inference on it as a service is not, without a deal. It’s a lighter restriction than many recent “open” licenses, but it’s a reminder that “open-weight” and “open-source” are not the same thing — read the license before you build a business on it.

Should You Self-Host in 2026?

A trillion-parameter MoE is not a laptop model — even with only ~52B active parameters, you’re looking at a serious multi-GPU deployment or a dedicated inference provider. So who is ML4 actually for?

1. Organizations with sovereignty requirements. Governments, defense, healthcare, and finance — anyone whose regulator (or threat model) says the data can’t leave the building. ML4 is now the strongest western option for that cohort.

2. High-volume agentic workloads. If you’re running thousands of coding or document-agent tasks a day, 47% fewer tokens per task can outweigh infrastructure costs — especially with $0.70/M API pricing as the fallback.

3. Builders fine-tuning for a niche. The weights plus Mistral’s Forge environment make ML4 a foundation for specialized derivatives. Expect domain-tuned variants to appear on Hugging Face within weeks.

For everyone else, the pragmatic move is unchanged: pick the tool that fits your workflow first, and let the model choice follow. If you’re evaluating coding agents, compare Mistral Large 4, Claude Code, and Cursor side by side — with pricing and alternatives for each — on aitrove.ai.

Frequently Asked Questions