Mistral launches Mistral Small 4: merging reasoning, multimodality and coding into one open-source model

On 16/03/2026, French AI company Mistral AI officially launched Mistral Small 4 — an open-source model under the Apache 2.0 license, free to self-host and customize. The most notable point: this is Mistral's first "Small"-tier model to merge three capabilities that were previously separate — reasoning, multimodality and a coding agent — into a single, unified model.

What is Mistral Small 4?

According to Mistral, Mistral Small 4 was released on 16/03/2026 under the Apache 2.0 license — meaning businesses can download it, self-host it on their own infrastructure and customize it without commercial restrictions. This is the core difference compared to closed models accessible only via API.

On architecture, this is a Mixture-of-Experts (MoE) model with 119 billion total parameters but only 6 billion active parameters per token (128 experts, 4 active at a time). This design gives the model the "knowledge" of a large model while keeping the compute cost per token close to that of a small model. The model supports a context of up to 256k tokens.

Merging three capabilities into one model

The standout point, per Mistral, is that Small 4 unifies three capability lines that were previously separate:

• Magistral — reasoning capability.
• Pixtral — multimodal capability (processing both images and text).
• Devstral — coding agent capability (reading, writing and editing code).

Instead of having to choose and operate multiple separate models for each need, users get a single model that handles all three. Mistral also added a reasoning_effort parameter, allowing you to tune the balance between reasoning depth and speed for each task — quick consideration or careful deliberation.

Performance and efficiency

According to Mistral, compared to the previous generation Mistral Small 3, Small 4 is up to 40% faster and achieves 3× the throughput. Mistral also says the model's output is about 20% shorter on the LiveCodeBench test — that is, more concise answers that still solve the problem, saving tokens.

These are all figures as published by Mistral; businesses should verify them on their own data and real tasks before moving to production.

Pricing and hardware requirements

On API cost, Mistral lists a price of around $0.10 per 1 million input tokens and $0.30 per 1 million output tokens — very competitive for a model that merges multiple capabilities.

Important for the self-hosting direction: according to Mistral, the model can run on around 4 NVIDIA H100 cards. Compared to frontier models that require large GPU clusters, this is a moderate hardware requirement, within reach of many businesses that want to run AI right inside their own infrastructure.

Why it matters

Mistral Small 4 reflects the trend of open-source models becoming increasingly powerful, compact and affordable: merging multiple capabilities, a compute-efficient MoE architecture, low API pricing and especially an Apache 2.0 license that gives businesses full freedom to self-host. When a model is good enough yet still runs on moderate hardware, the barriers to deploying AI right inside an organization drop significantly.

FAQ

Note: This article is compiled from publicly available sources as of 22/06/2026; the performance and hardware figures are as published by Mistral, and businesses should verify them independently. For reference only, not technical advice.