For Developers/Models/Mistral Small 4
Open SourceMistral AIReleased 2026-03

Mistral Small 4

Mistral's open-weight 119B MoE that handles chat, reasoning, image input and coding in one model

Context window

256K

Input / 1M tokens

$0.15

Output / 1M tokens

$0.60

Provider

Mistral AI

Mistral API, per 1M tokens: $0.15 input and $0.60 output; OpenRouter charges the same and lists cached input at $0.015. Open weights under Apache 2.0 are free to self-host. · Data verified 2026-11-11

Mistral Small 4 (March 16, 2026; API ID mistral-small-2603) is an Apache 2.0 open-weight Mixture-of-Experts model with 119B total parameters, 128 experts and 4 active per token, or about 6B active parameters (8B counting embedding and output layers). It combines instruct, reasoning, vision and agentic coding in a single model with a 256K context and text and image input, and a reasoning_effort parameter switches between fast replies ('none') and step-by-step reasoning ('high'). Mistral reports 40% lower end-to-end completion time and 3x more requests per second than Mistral Small 3. It is the only Mistral Small model still active on the Mistral API: Small 3.2 retired on July 31, 2026.

How to access it

Mistral API and Mistral AI Studio (model ID mistral-small-2603, labeled GA), NVIDIA build.nvidia.com and NIM, and OpenRouter (mistralai/mistral-small-2603). Open weights are on Hugging Face under Apache 2.0 for self-hosting with vLLM, SGLang, llama.cpp or Transformers.

Strengths

  • ✓Apache 2.0 open weights with only about 6B active parameters per token
  • ✓Low API price: $0.15 input / $0.60 output per 1M tokens
  • ✓One model for chat, reasoning, image input and coding, switched with reasoning_effort
  • ✓256K-token context
  • ✓Serving support in vLLM, SGLang, llama.cpp, Transformers and NVIDIA NIM

Best for developers who...

Self-hosted general assistantsCost-sensitive API workloadsAgentic coding and tool use on modest hardware

When to choose it (and when not to)

Reach for Mistral Small 4 when...

  • →You want an open-weight model you can run and fine-tune yourself
  • →You need cheap API calls with optional reasoning
  • →You want one model for text, image input and coding instead of several specialized ones

Look elsewhere if...

  • ✕You need frontier-level coding or reasoning (use Mistral Large 4 or a frontier model)
  • ✕You need more than 256K tokens of context
  • ✕You need audio input (use Voxtral)

How to use it

  • ›Set reasoning_effort to 'none' for quick chat and 'high' for multi-step problems
  • ›Use function calling and structured outputs for agent workflows
  • ›Send large offline jobs through the Batch API (/v1/batch)

Quickstart

Python
import os

from mistralai.client import Mistral

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

response = client.chat.complete(
    model="mistral-small-2603",
    messages=[
        {"role": "user", "content": "Write a Python function that validates an IBAN."},
    ],
)
print(response.choices[0].message.content)

Install the official SDK with pip install mistralai (or uv add mistralai). Pattern from Mistral's first-request quickstart.

API model id: mistral-small-2603

Benchmarks

BenchmarkScoreNotes
AA LCR0.72Mistral-reported, reasoning mode, with about 1.6K characters of output; Mistral says Qwen models need 3.5 to 4 times more output for similar scores
LiveCodeBenchBeats GPT-OSS 120BMistral-reported, reasoning mode, with 20% less output

Source: Mistral Small 4 announcement (Mistral AI)

Compare Mistral Small 4

Compare Mistral Small 4 with any other model

Build a comparison →
All model comparisons →

Learn the concepts