Open Source

Plot twist of the year: Mistral just re-licensed Mistral Large 3 under Apache 2.0. The same model that was their paid flagship six months ago is now free to deploy, modify, and ship commercially. Do not @ me — read the license file.

The Setup

Mistral Large 3 is a 123B dense model, agent-first, with tool-calling baked into the post-training. Function-calling reliability is the headline. It also speaks fluent JSON, which is a low bar that most models still trip over.

{`huggingface-cli download mistralai/Mistral-Large-3-Instruct \\
  --local-dir ./mistral-large-3

# minimum viable rig: 4x A100 80GB or 2x H100
vllm serve mistralai/Mistral-Large-3-Instruct --tensor-parallel-size 4`}

The Money Pattern

Tool calling is where this model earns its keep. Give it a tool schema and it'll plan a multi-step call, then actually execute the plan without going schizo on turn three.

{`tools = [{
    "type": "function",
    "function": {
        "name": "get_pipedrive_deal",
        "description": "Fetch a deal by id from Pipedrive",
        "parameters": {
            "type": "object",
            "properties": {"deal_id": {"type": "integer"}},
            "required": ["deal_id"],
        },
    },
}]

resp = client.chat.completions.create(
    model="mistralai/Mistral-Large-3-Instruct",
    messages=[{"role": "user", "content": "what's the status of deal 42?"}],
    tools=tools,
    tool_choice="auto",
)`}

The Catch

Still trails Claude Opus on long-form reasoning and creative writing. If your workload is "write me a 4000-word strategy memo with three competing angles," Opus still wins. If your workload is "call these eight tools in order and don't lie about the results," Large 3 ships.

The Verdict

For agent workflows on your own infra, this is the new pick. Better tool reliability than Llama 4 70B, more permissive than Qwen, and you don't have to phone home to a French API gateway. I'm wiring it into the FleetRoute orchestrator this week. The agent era is becoming open source by default.