Mistral AI has released Large 3, a 123 billion parameter model that the Paris-based company says is the first open-weight model with truly native agentic tool use. The model is available immediately through Mistral's API and as downloadable weights under the Mistral Research License.
Native Tool Calling
The headline feature is built-in function calling that does not rely on prompt engineering or system-level workarounds. Large 3 was trained from the ground up to understand tool definitions, generate structured function calls, interpret tool results, and chain multiple tool calls together to complete complex tasks.
In practice, this means developers can define tools using a standard JSON schema, and the model will reliably call them with correct parameters, handle errors, and retry with adjusted arguments when needed. Mistral reports a 94.2% success rate on the Berkeley Function Calling Benchmark — matching GPT-5 Turbo and surpassing Claude Sonnet.
"We trained tool use as a first-class capability, not a post-hoc addition," said Arthur Mensch, Mistral CEO. "The model understands tools the same way it understands language."
500K Context With Efficient Attention
Large 3 supports a 500,000-token context window using a sliding window attention mechanism that Mistral calls Grouped Sparse Attention. This allows the model to process long documents without the quadratic memory scaling that limits most transformer architectures.
In retrieval tests, the model maintains consistent accuracy across the full context window — a known weakness in many long-context models where information in the middle of the context is often missed. Mistral attributes this to a training approach that specifically targets uniform attention distribution.
Performance Benchmarks
On standard benchmarks, Large 3 positions itself as a strong competitor to closed-source models at a fraction of the cost:
- MMLU-Pro: 81.4% (vs GPT-5 at 86.1%)
- HumanEval: 89.2% (vs Claude Opus at 92.0%)
- Berkeley Function Calling: 94.2% (matching GPT-5 Turbo)
- Multilingual (European): Leads all models on French, German, Spanish, and Italian tasks
The multilingual strength is a continuation of Mistral's strategic advantage as a European AI company. Large 3 was trained with a significantly higher proportion of European language data than its American competitors.
Pricing and Access
Through Mistral's API, Large 3 is priced at $4 per million input tokens and $12 per million output tokens — roughly one-third the cost of GPT-5 Turbo for equivalent capabilities on tool-use tasks.
The open weights are available for download on Hugging Face under the Mistral Research License. Commercial use requires a separate agreement, though Mistral said it is introducing a simplified licensing process for startups with under $10 million in annual revenue.
Why It Matters
The open-weight AI space has consistently trailed closed-source models on agentic capabilities. Large 3 narrows that gap significantly. For developers building AI agents who need reliable tool calling without vendor lock-in, Mistral is offering a genuinely competitive alternative for the first time.
