I'm excited about this one - they seem to be directly targeting the "best model to run on a decent laptop" category, hence the comparison with Llama 3.3 70B and Qwen 2.5 32B.<p>I'm running it on a M2 64GB MacBook Pro now via Ollama and it's fast and appears to be very capable. This downloads 14GB of model weights:<p><pre><code> ollama run mistral-small:24b
</code></pre>
Then using my <a href="https://llm.datasette.io/" rel="nofollow">https://llm.datasette.io/</a> tool (so I can log my prompts to SQLite):<p><pre><code> llm install llm-ollama
llm -m mistral-small:24b "say hi"
</code></pre>
More notes here: <a href="https://simonwillison.net/2025/Jan/30/mistral-small-3/" rel="nofollow">https://simonwillison.net/2025/Jan/30/mistral-small-3/</a>
Note the announcement at the end, that they're moving away from the non-commercial only license used in some of their models in favour of Apache:<p><i>We’re renewing our commitment to using Apache 2.0 license for our general purpose models, as we progressively move away from MRL-licensed models</i>
Hi! I'm Tom, a machine learning engineer at the nonprofit research institute Epoch AI [0]. I've been working on building infrastructure to:<p>* run LLM evaluations systematically and at scale<p>* share the data with the public in a rigorous and transparent way<p>We use the UK government's Inspect [1] library to run the evaluations.<p>As soon as I saw this news on HN, I evaluated Mistral Small 3 on MATH [2] level 5 (hardest subset, 1,324 questions). I get an accuracy of 0.45 (± 0.011). We sample the LLM 8 times for each question, which lets us obtain less noisy estimates of mean accuracy, and measure the consistency of the LLM's answers. The 1,324*8=10,584 samples represent 8.5M tokens (2M in, 6.5M out).<p>You can see the full transcripts here in Inspect’s interactive interface: <a href="https://epoch.ai/inspect-viewer/484131e0/viewer?log_file=https%3A%2F%2Fepoch-benchmarks-production-public.s3.us-east-2.amazonaws.com%2Finspect_ai_logs%2FNbsnvBsMoMizozbPZY8LLb.eval" rel="nofollow">https://epoch.ai/inspect-viewer/484131e0/viewer?log_file=htt...</a><p>Note that MATH is a different benchmark from the MathInstruct [3] mentioned in the OP.<p>It's still early days for Epoch AI's benchmarking work. I'm developing a systematic database of evaluations run directly by us (so we can share the full details transparently), which we hope to release very soon.<p>[0]: <a href="https://epoch.ai/" rel="nofollow">https://epoch.ai/</a><p>[1]: <a href="https://github.com/UKGovernmentBEIS/inspect_ai">https://github.com/UKGovernmentBEIS/inspect_ai</a><p>[2]: <a href="https://arxiv.org/abs/2103.03874" rel="nofollow">https://arxiv.org/abs/2103.03874</a><p>[3]: <a href="https://huggingface.co/datasets/TIGER-Lab/MathInstruct" rel="nofollow">https://huggingface.co/datasets/TIGER-Lab/MathInstruct</a>
Not so subtle in function calling example[1]<p><pre><code> "role": "assistant",
"content": "---\n\nOpenAI is a FOR-profit company.",
</code></pre>
[1] <a href="https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501#function-calling" rel="nofollow">https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-...</a>
So the point of this release is<p>1) code + weights Apache 2.0 licensed (enough to run locally, enough to train, not enough to reproduce this version)<p>2) Low latency, meaning 11ms per token (so ~90 tokens/sec on 4xH100)<p>3) Performance, according to mistral, somewhere between Qwen 2.5 32B and Llama 3.3 70B, roughly equal with GPT4o-mini<p>4) ollama run mistral-small (14G download) 9 tokens/sec on the question "who is the president of the US?" (also to enjoy that the answer ISN'T orange idiot)
I tried just a few of the code generating prompts I have used last days, and it looks quite good and promising. It seems at least on par with qwen2.5-coder-32b which was the first local model i would actually use for code. I am also surprised how far we went with small models producing such more polished output in the last year.<p>On another note, I also wish they would follow up with a new version of the 8x7B mixtral. It was one of my favourite models, but at the time it could barely fit in my ram, and now that I have more ram it is rather outdated. But I don't complain, this model anyway is great and it is great that they are one of the companies which actually publish such models targeted to edge computing.
Finally, all the recent MoE model releases make me depressed with my mere 24GB VRAM.<p>> Note that Mistral Small 3 is neither trained with RL nor synthetic data<p>Not using synthetic data at all is a little strange
Interested to see what folks do with putting DeepSeek-style RL methods on top of this. The smaller Mistral models have always punched above their weight and been the best for fine-tuning.
Until today, no language model I've run locally on a 32GB M1 has been able to answer this question correctly: "What was Mary J Blige's first album?"<p>Today, a 4-bit quantized version of Mistral Small (14GB model size) answered correctly :)<p><a href="https://ollama.com/library/mistral-small:24b-instruct-2501-q4_K_M">https://ollama.com/library/mistral-small:24b-instruct-2501-q...</a>
This is really exciting—the 12-32b size range has my favorite model size on my home computer, and the mistrals have been historically great and embraced for various fine-tuning.<p>At 24b, I think this has a good chance of fitting on my more memory constrained work computer.
There's also a 22b model that I appreciate, since it _almost_ fits into my 12GB 3060. But, alas, I might need to get a new GPU if this trend of fatter smaller models continues.
Is there a good benchmark one can look at that shows the best performing llm in terms of instruction following or overall score?<p>The only ones I am aware of is benchmarks on Twitter, Chatbot Arena [1] and Aider benchmark [2]<p>1. <a href="https://huggingface.co/spaces/lmarena-ai/chatbot-arena-leaderboard" rel="nofollow">https://huggingface.co/spaces/lmarena-ai/chatbot-arena-leade...</a><p>2. <a href="https://aider.chat/docs/leaderboards" rel="nofollow">https://aider.chat/docs/leaderboards</a>
What's this stuff about the model catering to ‘80%’ of generative AI tasks? What model do they expect me to use for the other 20% of the time when my question needs reasoning smarts.
Used it a bit today on coding tasks and overall very pleasant. The combination of fast and fits into 24gb is also appreciated<p>Wouldn’t be surprised if this gets used a fair bit given open license
The AI race to zero continues to accelerate and Mistral has shown one card to just stay in the race. (And released for free)<p>OpenAI's reaction to DeepSeek looked more like cope and panic after they realized they're getting squeezed at their own game.<p>Notice how Google hasn't said anything with these announcements and didn't rush out a model nor did they do any price cuts? They are not in panic and have something up their sleeve.<p>I'd expect Google to release a new reasoning model that is competitive with DeepSeek and o1 (or matches o3). Would be even more interesting if they release it for free.
Curious how it actually compares to LLaMa.<p>Last year Mistral was garbage compared to LLaMa. I needed a permissive license, so I was forced to use Mistral, but I had LLaMa that I could compare it to. I was always extremely jealous of LLaMa since the Berkley Sterling finetune was so amazing.<p>I ended up giving up on the project because Mistral was so unusable.<p>My conspiracy was that there was some European patriotism that gave Mistral a bit more hype than was merited.