Mistral AI Python SDK

1.12.4 · active · verified Sat Feb 28

Official Python SDK for Mistral AI API. Has gone through two major breaking rewrites: v0 → v1 (MistralClient removed) and v1 → v2 (in pre-release on GitHub, not yet on PyPI). Most LLM-generated code references v0 patterns which no longer work.

Warnings

breaking MistralClient and MistralAsyncClient are removed. Use Mistral class for both sync and async.
Fix: Replace MistralClient(api_key=...) with Mistral(api_key=...). Same constructor signature.
breaking client.chat() call signature removed. Now client.chat.complete() for sync, client.chat.complete_async() for async.
Fix: Change client.chat(model=m, messages=msgs) to client.chat.complete(model=m, messages=msgs)
breaking Streaming chunk structure changed. chunk.choices[0] no longer works — must use chunk.data.choices[0].
Fix: Add .data between chunk and .choices in all streaming handlers
breaking ChatMessage class removed from mistralai.models.chat_completion. Use plain dicts or new typed message classes.
Fix: Replace ChatMessage(role='user', content='...') with {'role': 'user', 'content': '...'}
breaking tool_call_id is now mandatory in messages with role 'tool'. Omitting it causes API error.
Fix: Always include tool_call_id in tool response messages
gotcha pip install mistralai installs v1 (stable). GitHub main branch shows v2 docs which are NOT on PyPI yet. Don't follow GitHub README — it documents unreleased v2.
Fix: Use PyPI docs at pypi.org/project/mistralai for v1 reference. GitHub main = v2 pre-release.
gotcha Agents API requires Python 3.10+ and mistralai[agents] extra. Base install does not include it.
Fix: pip install 'mistralai[agents]' and ensure Python 3.10+
gotcha Azure and GCP deployments require separate packages: mistralai-azure and mistralai-gcp. Not included in base mistralai.
Fix: pip install mistralai-azure or pip install mistralai-gcp depending on deployment target

Install

pip install mistralai pip (installs v1 — stable)
uv add mistralai uv
pip install 'mistralai[agents]' pip (with agents support)
pip install mistralai-azure Azure deployment
pip install mistralai-gcp Google Cloud deployment

Imports

Mistral
```
from mistralai import Mistral
```
MistralClient and MistralAsyncClient were removed in v1. Single Mistral class handles both sync and async.
chat completion
```
client.chat.complete(model=..., messages=[...])
```
v0 used client.chat() directly. v1+ uses client.chat.complete(). Most LLM-generated code uses the old pattern.
streaming chunk
```
chunk.data.choices[0].delta.content
```
Streaming response structure changed in v1. chunk now wraps data — extra .data level required.
ChatMessage
```
{'role': 'user', 'content': '...'}
```
ChatMessage class removed. Use plain dicts or UserMessage/SystemMessage/AssistantMessage classes from mistralai.

Quickstart

Minimal chat completion using v1 SDK

import os
from mistralai import Mistral

client = Mistral(api_key=os.environ['MISTRAL_API_KEY'])

response = client.chat.complete(
    model='mistral-large-latest',
    messages=[{'role': 'user', 'content': 'Hello'}]
)
print(response.choices[0].message.content)

view raw JSON →