Mistral Small 3.1 Review: Benchmarks, Pricing & What's New
Published March 17, 2025 · Updated June 14, 2026
TL;DR
Mistral Small 3.1 (24B, Apache 2.0) was the best small open-source model when it launched in March 2025 — adding native multimodal vision and a 128k context window over its predecessor, while outperforming GPT-4o Mini and Gemma 3 on key benchmarks. It runs on a single RTX 4090 or 32GB Mac. If you are starting fresh in 2026, look at Mistral Small 4. If you are maintaining an existing deployment or need a lighter self-hosted model, Small 3.1 is still solid.
What Is Mistral Small 3.1?
Mistral Small 3.1 is a 24B parameter language model released by Mistral AI on March 17, 2025. It is an update to Mistral Small 3 with three major additions: native multimodal vision (image + text inputs), a quadrupled 128k token context window, and improved benchmark performance across text, multilingual, and long-context tasks.
Mistral describes it as "the best model in its weight class" at launch — and the benchmarks back that up: it outperforms both Gemma 3 and GPT-4o Mini on GPQA Diamond (graduate-level science) and instruction-following metrics, while delivering roughly 150 tokens per second on standard hardware.
What's New Over Mistral Small 3
Native multimodal vision
Mistral Small 3.1 is the first Small-series model to accept image inputs natively. You can send images alongside text in a single API call — no separate Pixtral model required. Use cases include document parsing, visual question answering, on-device image processing, visual inspection in quality control pipelines, and image-based customer support. The same weights handle both modalities.
128k token context window
Small 3 had a 32k context limit. Small 3.1 quadruples that to 128k tokens — enough for entire legal contracts, research papers, or multi-turn agent sessions spanning hours of conversation history. Long-context benchmarks show it outperforms competing models in the small-weight class.
Improved text and reasoning performance
On GPQA Diamond (graduate-level science questions) and standard instruction-following benchmarks, Mistral Small 3.1 beats GPT-4o Mini and Gemma 3. Mistral describes it as the best model in its weight class at launch. The improved reasoning capability also makes it a better base for community fine-tunes — models like DeepHermes 24B from Nous Research were built on the Small 3 base and saw strong results.
Apache 2.0 open weights
Like its predecessors, Mistral Small 3.1 is released under the Apache 2.0 license — arguably the most permissive license for any model at this capability level. You can download, self-host, fine-tune, and redistribute it commercially without Mistral's permission. Both the base pretrained checkpoint (Mistral-Small-3.1-24B-Base-2503) and the instruction-tuned version (Mistral-Small-3.1-24B-Instruct-2503) are on Hugging Face.
Consumer hardware deployment
At 24B parameters, Mistral Small 3.1 runs on a single NVIDIA RTX 4090 (24GB VRAM) or a Mac with 32GB unified memory. This is a meaningful practical threshold — it means research teams, indie developers, and privacy-conscious enterprises can self-host without renting GPU clusters. Supported frameworks include llama.cpp, vLLM, SGLang, and Hugging Face Transformers.
Key Benchmarks & Specs
| Benchmark / Spec | Small 3.1 | Notes |
|---|---|---|
| GPQA Diamond (graduate-level reasoning) | Beats GPT-4o Mini | Outperforms comparable proprietary models |
| Context window | 128k tokens | 4× larger than Mistral Small 3's 32k context |
| Inference speed | ~150 tok/s | Fast enough for real-time conversational applications |
| Multilingual | 40+ languages | Improved multilingual benchmark scores vs Small 3 |
| Hardware requirement | RTX 4090 / 32GB Mac | Runs on consumer or prosumer hardware |
| Model size | 24B parameters | Mistral-Small-3.1-24B-Instruct-2503 on Hugging Face |
Pricing & Availability
- Self-hosted (free): Download weights from Hugging Face under Apache 2.0 — no license fee, no usage restrictions. Instruct checkpoint and Base checkpoint are both available.
- Mistral La Plateforme API: Per-token pricing (check current rates). The Mistral Small tier includes Small 3.1 access.
- Third-party providers: Available on Together AI, Fireworks AI, Replicate, and other inference providers offering Mistral models.
- Le Chat: Accessible via Mistral's consumer product on free and Pro plans.
Pros & Cons
Pros
- ✓ Best-in-class for its size at launch — beats GPT-4o Mini, Gemma 3
- ✓ Apache 2.0: genuinely open, no commercial restrictions
- ✓ Native vision — no separate model needed for image tasks
- ✓ 128k context — 4× bigger than Small 3
- ✓ Runs on consumer hardware (RTX 4090, 32GB Mac)
- ✓ Both base and instruct checkpoints for fine-tuning
Cons
- ✗ Superseded by Mistral Small 4 (March 2026)
- ✗ Vision not as strong as GPT-4o on complex document tasks
- ✗ Still requires capable hardware — not edge/mobile deployable
- ✗ Mistral API capacity can lag OpenAI at peak demand
Who Should Use It
- Teams self-hosting a multimodal LLM — Small 3.1 is one of the few models that combines text + vision + open weights at a size that fits on a single GPU.
- Regulated industries (healthcare, finance, legal) needing on-premise deployment with a permissive license they can audit and modify.
- Researchers building fine-tuned models — the base checkpoint is an excellent starting point. The DeepHermes 24B reasoning model from Nous Research was built on Mistral Small 3 (base); Small 3.1 offers a stronger foundation.
- Cost-sensitive inference workloads — if GPT-4o Mini is too expensive at your volume and you want comparable or better quality, self-hosting Small 3.1 can cut costs significantly.
Frequently Asked Questions
What is Mistral Small 3.1?
Mistral Small 3.1 is a 24B parameter open-source language model released by Mistral AI on March 17, 2025. It adds multimodal vision and a 128k context window to the Mistral Small 3 base, while outperforming GPT-4o Mini and Gemma 3 on text and image benchmarks. It is released under the Apache 2.0 license.
How does Mistral Small 3.1 compare to GPT-4o Mini?
On GPQA Diamond and standard instruction-following benchmarks, Mistral Small 3.1 outperforms GPT-4o Mini. It also offers native multimodal vision, a 128k context window, and a fully open Apache 2.0 license — GPT-4o Mini is closed-source and proprietary. For cost-sensitive or self-hosted deployments, Small 3.1 is a strong alternative.
What hardware does Mistral Small 3.1 require?
Mistral Small 3.1 can run on a single NVIDIA RTX 4090 (24GB VRAM) or a Mac with 32GB unified memory. This makes it accessible for individual developers and small teams without GPU cluster access. For production inference at scale, providers like Mistral's La Plateforme API or third-party hosts (Together AI, Fireworks) offer managed access.
Is Mistral Small 3.1 free to use?
The model weights are free to download and self-host under the Apache 2.0 license, including for commercial use. Accessing it via Mistral's La Plateforme API incurs per-token costs. Le Chat (Mistral's consumer product) may offer free access with usage limits.
Does Mistral Small 3.1 support images?
Yes. Mistral Small 3.1 natively accepts both text and image inputs in the same API call. Vision capabilities cover image understanding, document parsing, visual Q&A, and multimodal chat. This is a new capability added over Mistral Small 3, which was text-only.
How does Mistral Small 3.1 relate to Mistral Small 4?
Mistral Small 3.1 was superseded by Mistral Small 4 in March 2026. Small 4 is a 119B MoE model that adds configurable reasoning, larger context (256k), and significantly higher throughput. If you are starting a new project in 2026, Small 4 or Medium 3.5 are the recommended options. Small 3.1 remains a solid choice for teams already running self-hosted deployments or who need a lighter weight footprint.
Explore Mistral Models on AISOTools
Compare Mistral Small 3.1 against newer models and alternatives.
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.