# Mistral Small 3 Mistral Small 3 is a 24B-parameter [[Large Language Models (LLMs)|LLM]] released by [[Mistral AI]] on January 30, 2025 (model ID `mistral-small-2501`). What made it notable at release: Mistral claimed it was **competitive with much larger models** like Llama 3.3 70B and Qwen 2.5 32B, while being more than 3x faster than Llama 3.3 70B on the same hardware. They positioned it as an open replacement for proprietary small models like GPT-4o mini. ## Why it mattered - **Apache 2.0 license**. Mistral moved away from its more restrictive Mistral Research License for this model. You can use it commercially, modify it, and run it wherever you want (see [[Apache 2.0 License]] and [[AI Open Weight Models]]) - **Runs locally**. Quantized, it's about a 14 GB download. That fits on a laptop with enough RAM or a single consumer GPU - **Cheap through the API**: $0.10 per million input tokens and $0.30 per million output tokens, half the price of the previous Mistral Small. GPT-4o mini was $0.15 / $0.60 at the time [[Simon Willison]] was excited about it for exactly that reason: models in the ~20 GB weight class hit a sweet spot for running on a laptop. ## Running it With [[Ollama]]: ```bash ollama run mistral-small:24b ``` With [[llm CLI|llm]] and the Ollama plugin: ```bash llm install llm-ollama llm -m mistral-small:24b "say hi" ``` Or through Mistral's API (La Plateforme) with the `llm-mistral` plugin: ```bash llm install -U llm-mistral llm -m mistral/mistral-small-latest "tell me a joke about a badger and a puffin" ``` It also works with [[LM Studio]] and llamafile, at multiple quantization levels (Q2 to Q8). ## What people reported The [[Hacker News]] thread (620 points) was mostly positive: - **Speed**: around 9-14 tokens/second on an M2 MacBook Pro with 64 GB, and around 46 tokens/second on an RTX 4090 with Q4 quantization - **Use cases**: document extraction and classification, email and support triage, intent detection, sentiment analysis, local RAG, function calling, and structured output - **Comparisons**: generally seen as a sweet spot between Llama 3.3 70B and Qwen 2.5 32B for local use, although some found Phi-4 better on their hardware - **"Open source" debate**: many argued "open weights" is the accurate term, since training data and code aren't released ## Successors Mistral Small 3 was the foundation for [[Mistral Small 3.1]] (March 2025), which added vision and a 128k context window, and later [[Mistral Small 4]]. ## References - Simon Willison's post: https://simonwillison.net/2025/Jan/30/mistral-small-3/ - Hacker News discussion: https://news.ycombinator.com/item?id=42877860 - Announcement: https://mistral.ai/news/mistral-small-3 ## Related - [[Mistral AI]] - [[Mistral Small 3.1]] - [[Mistral Small 4]] - [[AI Open Weight Models]] - [[Apache 2.0 License]] - [[Ollama]] - [[LM Studio]] - [[llm CLI]] - [[Simon Willison]] - [[Small Language Models (SLMs)]]