Mistral Small 3 Review: Benchmarks, Pricing & What's New
Published January 30, 2025 · Updated June 14, 2026
TL;DR
Mistral Small 3 (24B, Apache 2.0, January 2025) was the fastest open-source model in its weight class at launch — 150+ tok/s, 81%+ MMLU, and 3× faster than Llama 3.3 70B on the same GPU. It runs on a single RTX 4090 or 32GB Mac. It has since been superseded by Mistral Small 3.1 (vision + 128k context, March 2025) and Mistral Small 4 (March 2026). Small 3 remains a strong fine-tuning base for teams building reasoning models.
What Is Mistral Small 3?
Mistral Small 3 is a 24B parameter open-source language model released by Mistral AI on January 30, 2025. It is the first "Small" series model from Mistral to be optimized primarily for latency — designed to handle what Mistral calls the "80%" of generative AI tasks that require robust language and instruction following performance with very low latency.
The model is released under the Apache 2.0 license, with both a pretrained base checkpoint and an instruction-tuned version available on Hugging Face. It runs on a single NVIDIA RTX 4090 or a MacBook with 32GB of unified memory — a meaningful threshold that makes it accessible for individual developers, research teams, and privacy-conscious enterprises.
At launch, Mistral Small 3 was the most efficient model in its category: over 81% accuracy on MMLU, 150 tokens per second, and performance on par with Llama 3.3 70B and Qwen 32B — models three times its size. It was later updated with Mistral Small 3.1 in March 2025, which added native vision and a 128k context window.
Key Highlights
Latency-first architecture
Mistral Small 3 was specifically designed to be fast. Mistral achieved this by giving the model far fewer layers than competing 24B models, which substantially reduces the time per forward pass. The result: 150 tokens per second at over 81% MMLU accuracy — making it the most efficient model in its category at launch. Use cases where low latency matters — real-time chat, voice assistants, high-frequency function calling — are where Small 3 shines.
Competitive with models 3× its size
Mistral Small 3 is on par with Llama 3.3 70B instruct on instruction-following and general knowledge benchmarks, while being more than 3× faster on the same hardware. It also competes with Qwen 32B across coding, math, and language tasks. The efficiency advantage makes it an attractive replacement for teams currently running much heavier models.
Consumer hardware deployment
At 24B parameters, Mistral Small 3 runs on a single NVIDIA RTX 4090 (24GB VRAM) or a MacBook with 32GB unified memory. When quantized, the memory footprint drops further. This means individuals, small teams, and privacy-conscious enterprises can self-host without a GPU cluster. Supported frameworks include llama.cpp, vLLM, and Hugging Face Transformers.
Apache 2.0 — truly open weights
Mistral Small 3 is released under the Apache 2.0 license with both pretrained base and instruction-tuned checkpoints available on Hugging Face. You can download, fine-tune, redistribute, and use it commercially without restrictions — making it one of the most permissive capable models available at launch. This is the same license used by Mistral Small 3.1 and Mistral Small 4.
Clean base for fine-tuning
Mistral explicitly noted that Small 3 was not trained with RL or synthetic data — placing it earlier in the model production pipeline than models like DeepSeek R1. This makes it an excellent starting point for community fine-tunes, especially for reasoning capabilities. Projects like DeepHermes from Nous Research used the Small 3 weights as a base and demonstrated strong results from targeted RL training.
Key Benchmarks & Specs
| Benchmark / Spec | Mistral Small 3 | Notes |
|---|---|---|
| MMLU accuracy | 81%+ | Competitive with models 3× its size |
| Inference speed | ~150 tok/s | 3× faster than Llama 3.3 70B on same hardware |
| Model size | 24B parameters | Efficient for single-GPU deployment |
| Context window | 32k tokens | Extended to 128k in Small 3.1 (March 2025) |
| Hardware requirement | RTX 4090 / 32GB Mac | Runs on consumer or prosumer hardware |
| License | Apache 2.0 | Fully open — commercial use, fine-tuning, redistribution allowed |
Pricing & Availability
- Self-hosted (free): Download weights from Hugging Face under Apache 2.0 — no license fee, no usage restrictions. Both the base pretrained checkpoint and the instruction-tuned version (Mistral-Small-3-24B-Instruct-2501) are available.
- Mistral La Plateforme API: Per-token pricing via the Mistral Small tier. Check current rates — Small 3 may be accessed under the same pricing tier as Small 3.1.
- Third-party providers: Available on Together AI, Fireworks AI, and other inference providers that host open-weight models.
- Le Chat: Mistral's consumer product may offer access to Small 3 or its successors on free and Pro plans.
Mistral Small 3 vs. Small 3.1 vs. Small 4
Mistral has released three generations of the "Small" line since January 2025:
- Small 3 (January 2025): Text-only, 32k context, 24B dense model — fastest and lightest for pure text tasks.
- Small 3.1 (March 2025): Adds native vision (image inputs) and a 128k context window. Same 24B dense architecture with a broader capability set.
- Small 4 (March 2026): A 119B MoE model (6B active per token) that unifies reasoning, vision, and coding. 40% faster than Small 3, 256k context, configurable reasoning effort.
For new projects in 2025 or later, Small 3.1 or Small 4 are better starting points. Small 3 is most valuable as a fine-tuning base — particularly for teams building RL-based reasoning capabilities on a clean pretrained foundation.
Who Should Use Mistral Small 3?
- Teams fine-tuning an open-weight base for domain-specific applications (legal, medical, finance) who want a clean, non-RL pretrained starting point.
- Developers building high-throughput text pipelines where 150 tok/s matters and vision is not needed.
- Researchers who want to reproduce or extend community fine-tunes like DeepHermes that were built on the Small 3 weights.
- Organizations with data sovereignty requirements that need on-premise Apache 2.0 models.
Frequently Asked Questions
What is Mistral Small 3?
Mistral Small 3 is a 24B parameter open-source language model released by Mistral AI on January 30, 2025. It is optimized for low latency and runs at over 150 tokens per second — 3× faster than Llama 3.3 70B on identical hardware — while achieving 81%+ accuracy on MMLU. It is released under the Apache 2.0 license with both base and instruction-tuned checkpoints.
How does Mistral Small 3 compare to Llama 3.3 70B?
Mistral Small 3 is on par with Llama 3.3 70B on instruction-following benchmarks, but runs more than 3× faster on the same hardware. For teams where throughput and latency matter — real-time chat, function calling, agentic pipelines — Small 3 delivers comparable quality at a fraction of the inference cost.
What hardware does Mistral Small 3 require?
Mistral Small 3 runs on a single NVIDIA RTX 4090 (24GB VRAM) or a Mac with 32GB unified memory. When quantized (e.g., 4-bit), it fits in even less VRAM, making it accessible for hobbyists and developers. For production inference, Mistral's La Plateforme API and third-party providers like Together AI and Fireworks AI offer hosted access.
Is Mistral Small 3 still worth using in 2025–2026?
Mistral Small 3 was superseded by Mistral Small 3.1 in March 2025 (which adds vision and a 128k context window) and Mistral Small 4 in March 2026. For new projects, Small 3.1 or Small 4 are better choices. However, Small 3 remains a solid fine-tuning base for teams that specifically want a clean, non-RL pretrained model to build reasoning capabilities on top of.
Does Mistral Small 3 support images or multimodal inputs?
No — Mistral Small 3 is a text-only model. Multimodal support (image inputs) was added in Mistral Small 3.1, released March 2025. If you need vision capabilities, use Small 3.1 or Small 4.
What is the context window of Mistral Small 3?
Mistral Small 3 has a 32k token context window. This was quadrupled to 128k in Mistral Small 3.1 (March 2025) and further extended to 256k in Mistral Small 4 (March 2026). For long-document or long-context agentic tasks, the later models are more suitable.
Try Mistral Small 3
Access Mistral Small 3 via the official Mistral page or explore newer models in the series.
Affiliate disclosure: Some links on this page are affiliate links. If you sign up through them, AISO Tools may earn a commission at no extra cost to you. This never affects our rankings or reviews.
📬 Get the best new AI tools delivered weekly
One concise email with fresh launches, trending picks, and featured standouts.
Join thousands of professionals who discover the best AI tools every week. No spam — unsubscribe anytime.