Mistral AI shipped Mistral Small 3.1 on March 17, 2025. The Apache 2.0 licensed model has 24 billion parameters and a 128,000 token context window. It is available on Hugging Face and the Mistral AI homepage. This is not a research preview. The weights are free for commercial use and can be self-hosted. It is a direct challenge to closed models like GPT-4o mini and Claude Haiku that charge per million tokens. Mistral AI lists the model card and benchmark results. For a broader list of free models, see best free AI models in 2026.
Why release an open model now? Mistral Small 3.1 is aimed at developers who need low latency, low cost, and data control. It runs on one RTX 4090 or a single A100. You do not need an API key or a subscription. That is the key difference from closed offerings. The model performs at the level of OpenAI’s GPT-4o mini on general reasoning, but you can modify, fine-tune, and deploy it without vendor lock-in. This matters because major providers have tightened free tier limits and moved flagship models behind paywalls.
The release also matters for smaller labs and enterprises. A 24B model that beats a 27B Gemma 3 and a 32B Qwen 2.5 model on many tasks changes the math for on-device and edge AI. Mistral AI is not giving away a toy model. This is a production grade checkpoint with full weights and a permissive license. You can read more about how AI pricing changes in June 2026 are pushing developers toward open weights. The model card is on Hugging Face.
How Do the Top Options Compare?
| Model | Parameters | Context | License | MMLU | Self-host |
|---|---|---|---|---|---|
| Mistral Small 3.1 | 24B | 128k | Apache 2.0 | 81.5% | Yes |
| Mistral Large 2 | 123B | 128k | Mistral Research + Commercial | 84.0% | Yes (multi-GPU) |
| Codestral 25.01 | 22B | 256k | Non-Production + Commercial | N/A (code) | Yes |
| GPT-4o mini | Unknown | 128k | Proprietary | 82.0% | No |
Benchmarks are vendor reported values from Mistral AI and OpenAI. MMLU for Codestral is not a primary score because the model is specialized for code tasks. Always check the official model card for updated results.
1. Mistral Small 3.1 , Best open-weight model for single-GPU deployment
Mistral Small 3.1 is the latest Apache 2.0 open-weight release from Mistral AI. It packs 24 billion parameters and a 128,000 token context window into a checkpoint that fits on one RTX 4090 or A100 GPU. The model scores 81.5 percent on MMLU, 51.5 percent on GPQA, and 94.9 percent on HumanEval, which puts it ahead of Gemma 3 27B and Qwen 2.5 32B on several general reasoning tasks and just behind much larger closed models. Unlike API-only alternatives, you can download the weights from the Mistral AI homepage or Hugging Face and run them without per token fees. This matters for teams tired of AI free tier limits and surprise bills. The model also supports function calling and fine-tuning, so it can replace paid small models in production agents and RAG pipelines. If you are comparing open options, check the best free AI models list for how Mistral Small 3.1 fits against Llama and Qwen releases. The small size does not mean weak multilingual performance. Mistral tuned this version for Japanese, Korean, French, and German with improved tokenizer efficiency. You get stronger reasoning, a permissive license, and no vendor lock in. For developers burned by GitHub Copilot usage billing, self hosting Mistral Small 3.1 is one way to cut AI spend.
Key strengths:
- ✅ Full weights included for self-hosting
- ✅ Runs on a single 24GB GPU with quantization
- ✅ Apache 2.0 license allows commercial use and fine-tuning
- ✅ 128k token context supports long documents
- ✅ Beats Gemma 3 27B on MMLU and HumanEval
- ❌ Not as strong as Mistral Large 2 or frontier closed models
- ❌ Requires some GPU and model serving knowledge
- ❌ Benchmark edge over GPT-4o mini is narrow on some tasks
Who it’s for: Developers and startups that need a free, self-hosted small model for production without API costs.
2. Mistral Large 2 , Best open-weight alternative to GPT-4 class models
Mistral Large 2 is a 123 billion parameter open-weight model released under the Mistral Research License for non-commercial research and a separate commercial license. It has a 128,000 token context window and supports dozens of languages and code. On MMLU it scores 84.0 percent, which lands close to Llama 3.1 405B and GPT-4 class models but at a fraction of the hosting cost if you have the hardware. The catch is that a 123B model needs multiple GPUs or high end Mac setups to run at reasonable speed. The free access story is real for research, but commercial use requires a paid license. This is still important because it gives enterprises a path to audit and control a frontier grade model without sending data to a closed API. You can read about broader AI subscription tier changes to see why companies are looking at self hosting. The Mistral AI page has the full weights and license details. If you need a smaller open model for coding, compare with free AI coding tools. Mistral Large 2 excels at complex reasoning, long documents, and agentic workflows, but it is not a drop-in free replacement for every API. You need real infrastructure and ML engineering skill to serve it. For many teams, Mistral Small 3.1 is the practical free option, while Mistral Large 2 is a research asset or enterprise deployment target.
Key strengths:
- ✅ Strong MMLU score of 84.0 percent
- ✅ 123B parameters handle complex reasoning
- ✅ Multilingual support across many languages
- ✅ Full weights available for audit and control
- ✅ 128k context for long documents
- ❌ Commercial license is not free
- ❌ Requires multiple GPUs for reasonable speed
- ❌ Higher latency than smaller models
Who it’s for: Researchers and enterprises that need a frontier grade open weight model and can manage large GPU infrastructure.
3. Codestral 25.01 , Best open-weight code generation model from Mistral
Codestral 25.01 is a specialized open-weight code model from Mistral AI, released under the Mistral AI Non-Production License for research and testing but with a separate commercial license. It has 22 billion parameters and a 256,000 token context window, which makes it well suited for repository level code completion and agentic coding tasks. On HumanEval it scores 92.9 percent and on MBPP it scores 81.2 percent, beating many larger general models. The free access part applies to non-production use, which means you can download the weights and experiment locally, but you cannot deploy it in a paid product without a commercial agreement. That distinction matters as coding tools shift to usage-based billing and developers search for cheaper paths. You can find the model on Hugging Face and the Mistral AI homepage. Codestral 25.01 also works with popular IDEs through local servers, and it supports fill-in-the-middle. If you need a free coding assistant today, see the free AI coding tools landscape. The tradeoff is that Codestral is not a general chat model. It is tuned for code, so it will struggle with medical or legal text. For pure coding, it is one of the strongest small open weights, but the license prevents true open source commercialization unless you pay. That is an honest limitation local hackers should know before building a business on it.
Key strengths:
- ✅ 256k token context for repository level code
- ✅ 92.9 percent HumanEval score
- ✅ Strong code completion and fill-in-the-middle
- ✅ 22B parameters fit on one GPU
- ✅ Free for non-production research
- ❌ Non-commercial license for free use
- ❌ Not a general purpose chat model
- ❌ Commercial deployment requires a paid license
Who it’s for: Individual developers and researchers who need a free code model for experiments and non-production tools.
4. GPT-4o mini , Best closed small model for turnkey API access
GPT-4o mini is OpenAI’s low cost closed model. It is not open source and has no free self-host option, but it is easy to call from anywhere and does not require your own GPU. It scores 82.0 percent on MMLU and has a 128,000 token context window, but per token pricing adds up at scale. OpenAI has also tightened free tier limits and pricing throughout 2026, pushing developers to watch their spend. For a direct comparison of closed and open small models, see AI API free tier limits. GPT-4o mini supports vision, function calling, and JSON mode, which makes it convenient for quick projects. The downside is that you cannot inspect weights, cannot fine-tune on your own data for free, and you depend on OpenAI’s rate limits. Many teams use GPT-4o mini for prototyping then switch to Mistral Small 3.1 for production because the open weight model removes per token costs and data privacy risks. The OpenAI pricing page lists current rates. If you care about free and open access, GPT-4o mini is not the right choice long term. But if you need a reliable API without infrastructure, it still works well.
Key strengths:
- ✅ Easy managed API with no GPU setup
- ✅ Supports vision, function calling, and JSON mode
- ✅ Strong multilingual performance
- ✅ No infrastructure to maintain
- ❌ Closed weights with no self-hosting
- ❌ Per token fees add up at scale
- ❌ Rate limits and API dependency
Who it’s for: Teams that want a managed API for fast prototyping and can accept per token costs.
Frequently Asked Questions
Is Mistral Small 3.1 really free for commercial use?
Yes. Mistral Small 3.1 is released under the Apache 2.0 license, which permits commercial use, modification, and distribution without royalty payments. You still need your own hardware or a hosting provider, but the weights themselves are free. This is different from Mistral Large 2 and Codestral 25.01, which have separate commercial licenses.
Can I run Mistral Small 3.1 on a single GPU?
Yes. The 24 billion parameter model runs on a single RTX 4090, RTX 3090, or A100 with 24GB of VRAM using 4-bit or 8-bit quantization. Native FP16 requires about 48GB of memory, so a single A100 80GB works well. Quantized versions on one consumer GPU are common.
How does Mistral Small 3.1 compare to GPT-4o mini?
Mistral Small 3.1 scores 81.5 percent on MMLU compared to GPT-4o mini’s reported 82.0 percent. On HumanEval it scores 94.9 percent, which is much higher than GPT-4o mini’s code generation. The main advantage is self-hosting with no per token costs, but GPT-4o mini includes vision and a managed API.
What is the context window for Mistral Small 3.1?
The model supports a 128,000 token context window, which is enough for long documents and multi-step agent tasks. Codestral 25.01 doubles that to 256,000 tokens for repository level code.
Where can I download Mistral Small 3.1?
You can download the weights from the official Mistral AI homepage or from Hugging Face. The model card includes safetensors files, tokenizer, and a demo script. You do not need to request access or join a waitlist.
Does Mistral Small 3.1 support function calling and fine-tuning?
Yes. The model supports function calling and tool use, which makes it suitable for agentic applications. You can fine-tune it on your own data under the Apache 2.0 license and deploy the resulting model commercially.
What Should You Remember?
- Apache 2.0 license: Mistral Small 3.1 permits commercial use, modification, and redistribution with no per token fees.
- Single GPU: The 24B model runs on a RTX 4090 or A100 24GB with quantization.
- Benchmark edge: Mistral Small 3.1 scores 94.9 percent on HumanEval, beating GPT-4o mini on code.
- 128k context: Long document and agent tasks are supported without paid API limits.
- Closed models costly: GPT-4o mini still charges per token and has rate limits, while open weights remove vendor lock in.
- License variations: Mistral Large 2 and Codestral 25.01 are not fully open source for commercial use.
Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.
