Mistral Small 4 Review: Coding, Vision, and Reasoning in One Open Model
Mistral Small 4 is a 119B open-weight MoE trying to combine coding, vision, tools, and reasoning.
This review looks at the developer workflow upside, the deployment tradeoffs, and the benchmark caveats.
Mistral's pitch is not just another chat model. Small 4 is positioned as one model for instruction following, coding-agent workflows, image and document input, long context, structured outputs, tool calling, and reasoning effort you can turn on when a task needs it.
In this review, I break down what Small 4 actually is, why the 119B total parameter count and roughly 6.5B active parameters matter, where it is available, what OpenRouter lists for pricing, what self-hosting looks like with FP8 and NVFP4 paths, and why benchmark claims need careful scope. I also separate older Small 3 benchmark evidence from the Small 4 release so the verdict stays honest.
Chapters:
00:00 Small, but not tiny
00:20 The promise
00:46 What it actually is
01:10 Context, vision, and tools
01:41 Reasoning as a dial
02:06 Coding review
02:33 Vision and document work
02:56 Availability and cost
03:23 Self-hosting reality
03:52 Benchmark evidence
04:20 Older benchmark caveat
04:47 Counterpoints
05:13 Practical verdict
05:41 Test the whole workflow
Sources and references:
Mistral Small 4 launch: https://mistral.ai/fr/news/mistral-sm...
Mistral Small 4 model card: https://docs.mistral.ai/models/model-...
Mistral models overview: https://docs.mistral.ai/models/overview
Hugging Face model card: https://huggingface.co/mistralai/Mist...
Hugging Face Small 4 collection: https://huggingface.co/collections/mi...
OpenRouter listing: https://openrouter.ai/mistralai/mistr...
Artificial Analysis provider data: https://artificialanalysis.ai/models/...
SGLang deployment documentation: https://docs.sglang.io/cookbook/autor...
VentureBeat coverage: https://venturebeat.com/technology/mi...
NVIDIA Nemotron coalition announcement: https://nvidianews.nvidia.com/news/nv...
Mistral Small 24B Instruct 2501 model card: https://huggingface.co/mistralai/Mist...
Subscribe to Fluid Coding & AI for practical model reviews focused on developer workflows, coding agents, open models, and real deployment tradeoffs.
#MistralAI #OpenWeights #CodingAI #AIAgents #DeveloperTools