Mistral Small 3 - The NEW Mini Model Killer

Mistral has released “Mistral Small 3,” a new 24B-parameter open-weight model positioned as a fast, capable “workhorse” for everyday tasks—aimed at replacing frequent calls to smaller proprietary models rather than chasing the very top-end reasoning benchmarks. The pitch matters because it arrives amid market noise around reasoning-first systems, while many teams still need models that are quick, reliable, and cheap to run at scale (or locally) for high-volume workloads like chat, summarization, extraction, and tool use.

A key differentiator is licensing and deployment flexibility. Mistral Small 3 is released under the Apache 2 license, with both a base model and an instruct model published on Hugging Face. That open licensing enables users to modify the weights, fine-tune them, and serve them on-prem—plus it supports commercial and non-commercial use. The model ships with a 32k context window out of the gate, reducing the need for users to first fine-tune for longer inputs (though the transcript notes that community extensions to 64k or 128k are likely).

On capability, Mistral claims competitiveness against larger models for its size—specifically naming Llama 3.3 70B and Qwen 32B as reference points. The transcript frames the model as “not 7B,” but still small enough to be practical: it’s expected to be quantized, making it feasible to run on a laptop for private chat and local RAG without sending data to the cloud. At the same time, it’s also described as deployable in the cloud with low latency and high tokens-per-second throughput.

Language coverage is presented as broad but not fully “hardcore multilingual.” The model supports dozens of languages, with emphasis on Western European languages, plus Chinese, Japanese, and Korean. More importantly, Mistral Small 3 is built with agentic features in mind from the start—supporting native function calling and structured outputs (including formats like JSON-style extraction), which makes it well-suited for automation workflows rather than just free-form text.

In a hands-on test using a Colab setup and LangChain, the model produces well-formatted Markdown and generally thorough answers. It also demonstrates updated training data compared with older Mistral variants when asked about differences between models such as Llama, Vuna, and Alpaca. Persona adaptation works for typical role prompts, though at least one political figure example (“Kate the vice president” signing off as “Cala Harris”) lands incorrectly.

The most practical results come from instruction-following and tooling. When asked for succinct answers, it reliably returns short outputs (e.g., a one-word capital-of-England response), avoiding the tendency of some models to echo the question or pad with full sentences. For structured outputs and function calling, the transcript reports strong performance: tool calls are made with correct arguments for arithmetic examples, including multi-step calls. Overall, the model is framed as a promising foundation for fine-tuning—likely to spawn a wave of community variants—especially for local RAG and private assistants where cost and data control are central concerns.