Mistral AI Models

Mistral AI Models

Mistral AI produces highly efficient models that punch above their weight class.

Model Lineup

Open Models

API-Only Models

Key Innovations

Mixture of Experts (MoE)

• Uses 8 expert networks
• Only 2 experts active per token
• Near 8x7B quality at 7B inference cost

Sliding Window Attention

• Efficient long-context handling
• Reduced memory usage
• Better than naive attention

Running Locally

With Ollama

# Standard Mistral
ollama pull mistral

# Mixtral
ollama pull mixtral:8x7b

# Run
ollama run mistral

Python Integration

import ollama

response = ollama.chat(
    model='mistral',
    messages=[
        {'role': 'user', 'content': 'Write a haiku about coding'}
    ]
)
print(response['message']['content'])

Mistral API

Setup

from mistralai.client import MistralClient
from mistralai.models.chat_completion import ChatMessage

client = MistralClient(api_key="your-api-key")

response = client.chat(
    model="mistral-medium",
    messages=[ChatMessage(role="user", content="Hello!")]
)
print(response.choices[0].message.content)

Function Calling

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "parameters": {
            "type": "object",
            "properties": {
                "location": {"type": "string"}
            },
            "required": ["location"]
        }
    }
}]

response = client.chat(
    model="mistral-large-latest",
    messages=[ChatMessage(role="user", content="Weather in Paris?")],
    tools=tools
)

Comparison with Competitors

Best Practices

1. Use MoE for quality - Mixtral gives excellent results
2. Leverage context - 32K window is generous
3. Prompt engineering - Works well with chain-of-thought
4. Temperature - Default 0.7 works well