Mistral AI shipped Mistral Small 4 on May 20, 2026. The open-weight model includes 24 billion parameters, a 128,000 token context window, and native vision support. It is not a research preview. The weights are free to download, fine tune, and deploy under Apache 2.0. That license allows commercial use without royalties or sign-off. The release lands as developers search for open models that can read images, documents, and screenshots without paying per-token API bills. Mistral Small 4 is the first Small series model to combine vision and text in a single dense checkpoint. The model is available now.
Mistral AI announced the model on its official homepage and distributed weights through Hugging Face. The company did not paywall the full model behind an API or a custom non-commercial license. You can pull the safetensors, run the model with llama.cpp or vLLM, and host it on your own hardware. This matters because many free tier changes in 2026 have pushed developers toward self-hosted models. The wider free tier landscape has tightened, but Mistral chose the opposite path with a true Apache 2.0 release. That is not a minor detail. It means the model cannot be yanked by a pricing change.
Early benchmarks place Mistral Small 4 above Mistral Small 3.1 and competitive with small closed models. It scores 82.1 on MMLU, 62.3 on MATH, and 63.2 on MMMU. The vision module scores 84.9 on DocVQA and 78.5 on ChartQA. Those numbers mean the model can parse invoices, charts, and UI screenshots. It is not as strong as GPT-5.5 or Claude Opus on hard reasoning. But for local document work and lightweight agents, the gap is smaller than expected. The model runs on a single 24 GB GPU in 4-bit quantization, which opens the door for individuals and small teams.
The release matters because it gives open-source developers a multimodal model with no license friction. You can take the weights, modify them, and ship a product. There is no usage-based billing scheme, no credit pool, and no token counter. The AI price war has pushed API prices down, but those credits still expire. Mistral Small 4 changes the math for local inference. If you already run open models, this one removes the vision gap that forced many projects to call a paid vision API. That is the core shift.
How Do the Top Options Compare?
| Option | Best For | License | Vision | Context |
|---|---|---|---|---|
| Base Model | Self-hosting text | Apache 2.0 | No | 128k |
| Vision Variant | Multimodal local | Apache 2.0 | Yes | 128k |
| Le Chat Free | Zero-setup testing | Free tier | Yes | Limited |
| Ollama/llama.cpp | Consumer hardware | Apache 2.0 | Yes | 128k |
Quantized versions reduce quality slightly. Le Chat rate limits apply.
1. Mistral Small 4 Base Model , Best for self-hosting open weights
The base checkpoint is a 24 billion parameter dense transformer with a 128,000 token context window. It uses grouped query attention and rotary position embeddings. The model ships in BF16 safetensors and is ready for fine tuning with Axolotl, Transformers, and vLLM. You can find the weights through the Mistral AI homepage and Hugging Face. The Apache 2.0 license covers both the weights and the code, so you can use it in proprietary products without opening your own source.
Benchmark results show the model outperforms Mistral Small 3.1 by 4.1 points on MMLU and 7.8 points on GPQA. That is meaningful for a dense model that does not rely on mixture-of-experts routing. The context window handles long PDFs, transcripts, and code repos. Early users report stable attention beyond 64k tokens, with acceptable quality at 128k. The base model does not include the vision tower, but the weights are still useful for text-only pipelines that need low VRAM.
If you compare it with closed models, Mistral Small 4 is not trying to beat GPT-5.5 on math. It is positioned as a small, free, open model that handles real business documents. The open-source LLM guide lists it as a strong pick for local coding and agent work. The lack of a usage fee is the main advantage.
Key strengths:
- ✅ Full Apache 2.0 commercial license
- ✅ 24B dense model with 128k context
- ✅ Outperforms Mistral Small 3.1 on MMLU and GPQA
- ✅ Runs in 4-bit quantization on a 16 GB GPU
- ✅ No API key or token cost
- ❌ Base checkpoint lacks vision, so you need the vision variant for images
- ❌ BF16 weights require 48 GB disk and high RAM for full fine tuning
- ❌ Smaller than frontier open models like DeepSeek V4
Who it’s for: Developers who want a free, self-hosted text model with strong local inference.
2. Mistral Small 4 Vision , Best for multimodal local pipelines
The vision version of Mistral Small 4 adds a vision encoder that converts images into tokens the language model can read. It accepts images at up to 1024x1024 resolution and supports JPEG, PNG, and WebP. The model scores 84.9 on DocVQA and 78.5 on ChartQA, which shows it can read documents, charts, and UI screenshots. That makes it useful for invoice parsing, receipt classification, and browser automation. You can run the model on a single 24 GB GPU in 4-bit quantization.
The vision module does not require a separate paid OCR service. You can feed a screenshot of an error and ask the model to explain the stack trace. You can upload a chart and ask for an executive summary. The open-source multimodal space has grown in 2026, but many options are not truly open. Mistral Small 4 stands out because both the text and vision weights are Apache 2.0. There is no awkward research-only restriction on the image encoder.
You can load the vision model through Hugging Face transformers or use llama.cpp with the mmproj file. The recommended path is to download the safetensors and the vision projector, then run a server. Early tests show the model can process a 100 page PDF by splitting pages into images, though the 128k token context may limit how many images fit at once. For most real workflows, 20 to 30 pages per request is the practical ceiling.
Key strengths:
- ✅ Native image understanding with no paid OCR
- ✅ Apache 2.0 license covers vision and text
- ✅ Strong DocVQA and ChartQA results
- ✅ Works with llama.cpp and Transformers
- ✅ No image-specific API fees
- ❌ Vision variant is larger on disk due to the vision encoder
- ❌ High resolution images consume many tokens
- ❌ MMMU score still trails top closed multimodal models
Who it’s for: Teams that need local document parsing, screenshot analysis, or chart summarization.
3. Mistral Small 4 via Le Chat Free Tier , Best for zero-setup testing
Mistral has made Mistral Small 4 available through Le Chat, its free chat interface. You can upload an image, ask questions, and test the model without installing anything. The Le Chat free tier includes limited daily messages, but it is enough to evaluate the model. You do not need a credit card or an API key. That is useful when you want to know if the vision quality fits your use case before downloading 14 GB of weights.
The free tier is not intended for production. Rate limits apply, and the service may route some requests to a smaller model during peak load. Mistral does not publish exact token limits for the free chat, but users have reported a 5 hour reset pattern similar to other free tiers. The free tier shift across the industry makes this worth watching. If you need guaranteed throughput, use the open weights or the paid API.
Still, Le Chat is the fastest way to test Mistral Small 4 vision. You can drag a screenshot of a spreadsheet and ask for a formula. You can upload a contract image and ask for red flags. The interface is clean and the model responds quickly. For developers who are tired of API credits expiring, the free chat is a low pressure starting point.
Key strengths:
- ✅ No setup, no GPU, no API key
- ✅ Free instant access through Le Chat
- ✅ Good for quick vision tests
- ✅ No credit card required
- ❌ Daily rate limits apply
- ❌ Production use requires API or self-hosting
- ❌ May route to smaller model during load
Who it’s for: Developers and tinkerers who want to try the model before downloading it.
4. Mistral Small 4 via Ollama or llama.cpp , Best for consumer hardware
If you have a gaming GPU or a Apple Silicon Mac, you can run Mistral Small 4 locally. The 4-bit quantized version is about 14 GB. A 16 GB MacBook with M2 or M3 can run the text model at 8 to 12 tokens per second. A 24 GB RTX 4090 can run the vision model comfortably. llama.cpp latest releases added support for multimodal GGUF with mmproj, so you can load the vision variant without a Python environment.
The setup is not one click. You need to download the GGUF and the vision projector, then start a local server or use Ollama. The local model guide covers common pitfalls like KV cache size and context truncation. For Windows users, LM Studio and Ollama expose a chat UI. For Linux users, llama.cpp server is the most flexible option. Once running, the model has no network calls and no data leaves your machine.
Local inference gives you fixed costs. You pay for electricity and hardware, not tokens. The downside is that the model is slower than an API. On a 24 GB GPU, you can expect 20 to 30 tokens per second in 4-bit. On CPU, it drops to 1 to 3 tokens per second, which is too slow for interactive use. The open-source self-host guide recommends 16 GB of unified memory as the practical minimum.
Key strengths:
- ✅ Runs offline with no data leakage
- ✅ Fixed hardware cost, no per-token fees
- ✅ Works on Apple Silicon and RTX GPUs
- ✅ Full control over model and context
- ❌ Quantization lowers some benchmark accuracy
- ❌ Vision variant needs more VRAM
- ❌ Setup requires technical skill
Who it’s for: Privacy-focused developers and homelab users with a 16 GB GPU or Mac.
Frequently Asked Questions
Is Mistral Small 4 really free for commercial use?
Yes. The weights and code are released under Apache 2.0. You can use the model in commercial applications, modify it, and do not need to open source your own project. Royalties and license fees do not apply.
How much VRAM do I need to run Mistral Small 4?
The 4-bit quantized text model needs about 14 GB of memory and runs on a 16 GB MacBook or a 16 GB GPU. The vision variant and longer context require more, but a 24 GB GPU is comfortable. Full BF16 inference needs 48 GB.
Does Mistral Small 4 support image inputs?
Yes, the vision variant supports image inputs at up to 1024x1024. It can read documents, charts, screenshots, and photos. The base model is text only, so download the vision checkpoint if you need multimodal work.
How does Mistral Small 4 compare to Llama 4 Scout?
Mistral Small 4 is smaller and less powerful on complex reasoning than Llama 4 Scout. It wins on license simplicity because Llama models have community restrictions. For local document and vision tasks, Mistral Small 4 is often easier to deploy and cheaper to run.
Where can I download Mistral Small 4?
You can find the safetensors and GGUF files on Hugging Face and through the official Mistral AI homepage. The company has not published a GitHub repo for the model itself, but runtimes like llama.cpp and vLLM support it.
What is the context window?
The context window is 128,000 tokens. That allows long PDFs, transcripts, and multi-file code sessions. Memory use increases with context, so local users may want to cap the context at 32k for speed.
What Should You Remember?
- Apache 2.0 license means free commercial use with no royalties.
- 24B dense model with 128k context outperforms Mistral Small 3.1.
- Native vision support reads documents, charts, and screenshots locally.
- Hugging Face and Mistral AI host the weights; download and run.
- Le Chat free tier offers zero-setup testing with rate limits.
- Local deployment on 16 GB hardware is possible but requires quantization.
- Benchmarks show strengths in DocVQA and MMLU but not frontier math.
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.
