Krux
March 20, 2026
Mistral Launches Enterprise AI Training System With 119B Model
Published: March 20, 2026 at 12:58 AM
Updated: March 20, 2026 at 12:58 AM
100-word summary
Mistral just launched Forge, a system that lets companies train frontier-grade AI models on their own data without sending it to OpenAI or Anthropic. The platform handles everything from initial training to ongoing improvement via reinforcement learning. It ships with Mistral Small 4, a 119-billion-parameter model that activates just 6.5 billion parameters per request by routing through 4 of its 128 expert networks. That architecture means faster responses without sacrificing capability. The model handles 256,000-token contexts and processes both text and images. Forge targets government agencies and financial firms that can't risk data leaks. The bet: enterprises will pay to own the entire training pipeline once AI agents start handling...
What happened
Mistral just launched Forge, a system that lets companies train frontier-grade AI models on their own data without sending it to OpenAI or Anthropic. The platform handles everything from initial training to ongoing improvement via reinforcement learning. It ships with Mistral Small 4, a 119-billion-parameter model that activates just 6.5 billion parameters per request by routing through 4 of its 128 expert networks. That architecture means faster responses without sacrificing capability. The model handles 256,000-token contexts and processes both text and images.
Why it matters
Forge targets government agencies and financial firms that can't risk data leaks. The bet: enterprises will pay to own the entire training pipeline once AI agents start handling sensitive internal workflows.