What makes Mistral's mixture-of-experts architecture different from a dense model, and when should I choose one over the other for production?

The core distinction is how parameters are activated per token:

• Dense models use 100% of their parameters for every token. Computation scales with total parameter count.
• Mixture-of-Experts (MoE) models route each token to a small subset of "expert" networks. Mixtral 8×7B activates only ~13B of its 47B parameters per token.

This gives you the quality of a larger model at the inference cost of a smaller one:

from mistralai import Mistral

client = Mistral(api_key="...")
response = client.chat.complete(
    model="mistral-large-latest",
    messages=[{"role": "user", "content": "Explain MoE"}]
)

Choose MoE when you want high quality with controlled latency. Choose dense when you need predictable per-token cost and simpler deployment.

Can you show me how to stream the response with function calling?

Yes — here's a streaming example with a tool definition. The stream yields chunks you can aggregate as they arrive:

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get current weather",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"]
        }
    }
}]

stream = client.chat.stream(
    model="mistral-large-latest",
    messages=messages,
    tools=tools
)
for chunk in stream:
    print(chunk.data.choices[0].delta.content, end="")

The model emits a tool_calls object when it decides to call a function — you execute it locally and feed the result back.