Skip to main content

Mistral's 1T-Parameter Giant Goes Open Source—with a Catch

Mistral's 1T-Parameter Giant Goes Open Source—with a Catch

French AI company Mistral just did something that would make most closed-source giants sweat: it released the full weights of its flagship model, Mistral Large 4 (code-named le Chonk), for anyone to download, modify, and even use commercially. The move, announced on October 29, puts a 1-trillion-parameter model into the hands of developers and enterprises worldwide.

Image

The Fine Print: Free Weights, Paid Resale

The license is a modified MIT—generous for personal, academic, and commercial use. But there's a catch: if you want to act as a cloud provider selling compute and resell the model as a service (MaaS), you'll need to negotiate a separate deal with Mistral. That "free weights, gatekeeping on resale" approach is Mistral's signature move, setting it apart from the likes of OpenAI and Google.

A Trillion Parameters, Built on European Soil

Mistral Large 4 is a native multimodal model with 1 trillion total parameters and 49 billion active parameters (52 billion including embedding and output layers). It was trained from scratch entirely in Mistral's own European data centers, using 3,800 NVIDIA Grace Blackwell GPUs. This version has been optimized and upgraded after real-world testing by governments and enterprises over the past three weeks.

Image

Token Efficiency: The Real Selling Point

What's the pitch to developers? It's not just raw power—it's cost. On the Artificial Analysis Coding Agent Index, ML4 scores 48%, roughly on par with leading open-source model GLM-5.3. But here's the kicker: it consumes 47% fewer tokens per task on average. That means doing the same work for less money. Online pricing is set at $0.70 per million input tokens, $0.07 for cached input, and $2.10 for output. Mistral is clearly positioning this as "same capability, lower invocation cost."

Key Points

  • Mistral Large 4 weights are now open source under a modified MIT license.
  • 1 trillion total parameters, 49 billion active; trained on 3,800 NVIDIA GPUs in Europe.
  • License allows free commercial use, but MaaS resale requires a separate deal.
  • Token efficiency is a major selling point: 47% fewer tokens per task than competitors.
  • Pricing starts at $0.70 per million input tokens, undercutting many rivals.