Mistral just did something most AI companies avoid. Instead of releasing three separate specialized models and making developers juggle between them, they merged everything into one.
Mistral Small 4 combines reasoning, multimodal and agentic coding into a single open source model. Until today if you wanted Mistral’s best reasoning you used Magistral. Best coding agents you used Devstral. Image and document understanding you used Pixtral. Three different models, three different integrations & three different things to maintain.
Now it is one model. Apache 2.0 licensed & Available on huggingface.
It has 119 billion total parameters but only 6 billion active at any time. That efficiency gap is what makes it practical to actually deploy.
If you have been waiting for an open source model that does not force you to choose between speed, reasoning and vision, this is worth paying attention to.
Table of contents
What is Mistral Small 4
Mistral Small 4 is a multimodal AI model that accepts both text and image inputs and generates text outputs. It handles general chat, coding tasks, document analysis, complex reasoning and visual understanding all in one place.
The architecture is Mixture of Experts, 128 experts total with only 4 active per token. That is how you get 119 billion total parameters behaving like a 6 billion parameter model at inference time. More capability, lower compute cost.
Context window sits at 256K tokens. Compared to some models offering 1 million tokens or more that is on the lower end but for the vast majority of real world use cases — long documents, extended conversations, large codebases, 256K is more than enough considering its capabilities.
What makes it different from most models its size is the reasoning_effort parameter. You can tell it how hard to think. Set it low and you get fast responses for simple tasks. Set it high and you get deep step by step reasoning for complex problems. Same model, different modes, no switching required.
It is designed for three types of users. Developers who need coding automation and agentic workflows. Enterprises who need document understanding and chat assistants. Researchers who need reliable math and reasoning capabilities.
The “Three-in-One” Model Strategy
Until today if you wanted the best of Mistral you needed three different models.
Magistral for complex reasoning and research tasks. Devstral for agentic coding, it held the top open source spot on SWE-bench Verified at 46.8% and was specifically built with All Hands AI for software engineering workflows. Pixtral for vision and multimodal tasks like images, documents, charts, visual analysis.
Mistral Small 4 replaces all three. The reasoning capabilities of Magistral are now configurable on demand. The agentic coding performance of Devstral is built in. The visual understanding of Pixtral is native. One model, one integration, one deployment.
For developers building applications that need more than one of these capabilities this is genuinely significant. Instead of routing requests between specialized models or maintaining separate infrastructure for each use case you have one model that adapts to whatever the task requires.
That is not a minor convenience improvement. For teams running multiple Mistral models in production it simplifies the entire stack.
How it performs
The benchmark that stands out most is LiveCodeBench. Mistral Small 4 with reasoning beats GPT-OSS 120B while producing 20% shorter outputs. That second part matters as much as the first — shorter outputs mean lower latency, lower cost and a better experience for anyone using it in a product.
On AA-LCR long context reasoning it scores 0.72 with just 1.6K characters of output. Qwen models need 3.5 to 4 times more output to reach comparable performance on the same benchmark. Again the efficiency story is as interesting as the raw score.
On AIME 2025 math reasoning it is competitive with models significantly larger than its 6B active parameter count.
Honest caveat — these benchmarks are from Mistral’s own evaluation pipeline. Independent third party benchmarks will give a fuller picture once the community has had time to test it properly. That said the efficiency numbers are verifiable and the pattern of doing more with less output is consistent across multiple benchmarks.
For a model that runs on 4x H100s at minimum the performance per active parameter is genuinely impressive.
Also Read: Reka Edge: The 7B Multimodal AI Model That Beats Gemini 3 Pro on Object Detection
How to Run Mistral Small 4 Locally
The easiest way to get started locally is LM Studio. It is available on their site, search for Mistral Small 4 in the model section, download and run.
For hardware you will need a capable GPU to run it comfortably locally. Given the 119B total parameters with 6B active the rough requirement is around 24GB VRAM minimum for smooth inference. More is better.
If your hardware is not there yet Mistral provides an API with full documentation. Straightforward to set up and the model is available immediately without any local infrastructure.
For developers who want more control it is also available on vLLM, SGLang, llama.cpp and Transformers. Pick whichever fits your existing workflow.
Three Specialized Models, One Unified System
Mistral Small 4 is not just a new model release. It is a product decision that simplifies how developers work with Mistral’s ecosystem. Instead of maintaining three separate integrations for reasoning, coding and vision you have one model that handles all three with a single deployment.
For anyone already using Magistral, Devstral or Pixtral separately this is worth evaluating seriously. The consolidation alone reduces complexity even before you factor in the performance improvements.
If you want to explore more open source AI models check out the AI Models section.
