Mistral Small 4, one model for reasoning, visioning, and coding

---

Introduction

Managing multiple specialized models within a single AI pipeline adds deployment complexity and multiplies infrastructure costs. Mistral AI announced on March 16, 2026 the launch of Mistral Small 4, a single model that absorbs the roles previously held by Magistral (reasoning), Pixtral (vision), and Devstral (code). For an enterprise AI admin, the message is straightforward : one API integration, three specialties covered.

---

Architecture and capabilities

Mistral Small 4 is built on a Mixture of Experts (MoE) architecture with 119 billion total parameters, of which approximately 6 billion are activated per token. The model uses 128 experts with 4 active simultaneously, keeping the computational footprint at inference low despite the model’s raw size.

Key characteristics :

• Context window : 256,000 tokens, sufficient to process long documents or extended conversation histories.
• Native multimodality : Text and image as input, text as output. Small 3 was strictly text-only; that limitation is gone.
• Configurable reasoning : The new reasoning_effort parameter accepts the values none or high, making it possible to switch between a fast response and deep reasoning based on the use case, without changing models.
• Language support : 24 languages including French, English, Spanish, German, Chinese, Japanese, and Arabic.

---

What changes compared to Small 3

The most significant difference from Mistral Small 3 is not purely about performance, it is the consolidation of several models into a single deployment point. Previously, a complete AI stack required routing requests to distinct models depending on the task, a vision model like the one introduced in Mistral OCR, a code model like Mistral Code. Small 4 eliminates that routing.

On raw performance, Mistral claims a 40% reduction in end-to-end latency and three times higher throughput in requests per second compared to Small 3. These figures are published by Mistral without a detailed methodology; they should be treated as indicative until an official technical report is available. Independent measurements by Artificial Analysis place the model at 171.8 tokens per second with a time-to-first-token (TTFT) of 0.76 seconds via the Mistral API.

---

Benchmarks

In the absence of an official technical report at the time of the announcement, the results below come from secondary sources and Mistral publications. They provide a positioning indication but do not constitute a complete, independent evaluation.

	Mistral Small 4	GPT-4o-mini
GPQA Diamond	71.2%	40.2%
MMLU-Pro	78.0%	64.8%
LiveCodeBench	Outperforms GPT-OSS 120B (−20% tokens)	-

On the AA LCR (weighted response length) benchmark, Small 4 scores 0.72 for 1,600 characters, against ranges of 5800 to 6100 characters for Qwen 3.5-122B, reflecting a notably concise output style.

---

Access and deployment

Mistral Small 4 is available under the Apache 2.0 license, authorizing commercial use without revenue restrictions, unlike Meta’s LLaMA conditions, which impose usage thresholds based on organization size.

Available access channels at launch:

• Mistral API and AI Studio : Pricing sourced via Artificial Analysis and OpenRouter: $0.15 per million input tokens, $0.60 per million output tokens, not yet published on the official Mistral pricing page at time of writing.
• Hugging Face : mistralai/Mistral-Small-4-119B-2603
• NVIDIA NIM, vLLM, llama.cpp, SGLang, Transformers, Axolotl (fine-tuning)

For self-hosting, the minimum hardware requirement is significant:

• 4× NVIDIA HGX H100
• 2× HGX H200
• 1× DGX B200

This level of infrastructure rules out consumer GPU deployments or lightweight on-premises setups; the API remains the most accessible option for most organizations.

---

Conclusion

Mistral Small 4 addresses a concrete enterprise AI stack problem, instead of maintaining three separate models for reasoning, vision, and code, a single deployment covers all three. The Apache 2.0 license, the reasoning_effort parameter, and the 256k context make it a serious candidate for diverse automation workflows. The absence of an official technical report is a reason to validate benchmarks against internal use cases before any production commitment.

---

Sources

Mistral AI - Mistral Small 4

Test Le Chat by Mistral AI

Hugging Face - mistralai/Mistral-Small-4-119B-2603

Artificial Analysis - Mistral Small 4

OpenRouter - mistralai/mistral-small-2603

MindStudio - What is Mistral Small 4

Emelia - Mistral Small 4 Complete Guide

---

Did you enjoy this post ? If you have any questions, comments or suggestions, please feel free to send me a message from the contact form.

Don’t forget to follow us and share this post.

---