Mistral AI Open-Source Release: Performance and Free Access
Quick Answer: Mistral Small 4, released March 16, 2026 under the Apache 2.0 license, is a 119B-parameter mixture-of-experts model that combines text generation, reasoning, vision, and coding in a single download. It is free to self-host via Hugging Face or Ollama, and costs $0.15 per million input tokens via API.
For developers who have been juggling multiple Mistral models – one for fast chat, another for reasoning, a third for coding agents – Mistral Small 4 is a significant quality-of-life upgrade. Released on March 16, 2026, it collapses three separate model families (Magistral for reasoning, Devstral for coding, and the standard Small instruct line) into one Apache 2.0-licensed download. More importantly for cost-conscious users, it does so at a price point substantially below the frontier closed models it competes with on many benchmarks.
This guide covers the full technical picture, the honest performance story, and every route to using Mistral Small 4 for free or near-free today. We also look at the broader Mistral 3 family released in December 2025, including Mistral Large 3, which brought frontier-class open weights to developers for the first time.
What Is Mistral Small 4 and What Makes It Different?
Mistral Small 4 is a mixture-of-experts (MoE) language model from the French AI company Mistral AI. Unlike a dense model where every parameter activates for every token, Small 4 uses 128 experts but only routes each token through 4 of them. That means 119 billion total parameters, but only around 6.5 billion active per token during inference. The practical result: memory and compute requirements that are far closer to a 7-14B dense model than a 100B+ monolith.
What sets Small 4 apart from previous Mistral releases is unification. Prior to this model, developers who needed fast chat responses used Mistral Small 3.2, those who needed complex multi-step reasoning used the Magistral series, and those building coding agents relied on Devstral. Small 4 replaces all three. You toggle the behavior at the API level with a single reasoning_effort parameter rather than spinning up different endpoints.
What Are the Key Technical Specs of Mistral Small 4?
119B Total Parameters
6.5B Active per Token
256K Context Window
128 MoE Experts (4 active)
40% Faster Completion
3x Throughput vs Small 3
The full architecture breakdown from Hugging Face includes:
Architecture – Sparse Mixture of Experts (MoE) with 128 experts, 4 active per token
Parameters – 119B total, approximately 6.5B activated per forward pass (8B including embedding and output layers)
Context window – 256,000 tokens, suitable for processing entire codebases or long documents
Multimodal inputs – Native text and image input, text output; useful for document parsing and visual analysis
Reasoning modes – Configurable per request: fast instruct mode or deep chain-of-thought reasoning
Multilingual – Strong support for English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, and Arabic
License – Apache 2.0, permitting commercial use, fine-tuning, and redistribution with no royalties
Mistral also released an NVFP4-quantized checkpoint (Mistral-Small-4-119B-2603-NVFP4) and a speculative decoding Eagle head for further inference optimization. Both are available on Hugging Face alongside the standard weights.
How Does Mistral Small 4 Perform Against Other Models?
Performance is where the community reaction gets nuanced. On coding benchmarks such as HumanEval, Mistral Small 4 sits in the same performance tier as Claude Haiku 3.5 and Qwen 2.5 14B – solid but not leading. On instruction following and multilingual tasks, community testing consistently puts it above models twice its active-parameter count.
The throughput story is more impressive. In latency-optimized deployments, Small 4 achieves a 40% reduction in end-to-end completion time versus Mistral Small 3. In throughput-optimized deployments, it handles 3x more requests per second than its predecessor. For production workloads where you are serving thousands of concurrent users, this translates directly to infrastructure cost savings.
Honest benchmark caveat: Several community benchmarks published in late April 2026 noted that Small 4’s flagship scores are not best-in-class on any single standard benchmark when compared to the newest closed models from OpenAI and Anthropic. Its advantage is the combination of open weights, multimodality, and reasoning toggleability at a cost 5 to 7 times lower than comparable closed-model API pricing.
For more context on how open-source models stack up against paid alternatives across your specific use case, see our AI model comparison tool.
How Can You Use Mistral Small 4 for Free in 2026?
Because Small 4 is Apache 2.0, “free” has a meaningful definition here. There are three main routes:
Mistral Le Chat (free web tier) – Mistral’s own chat interface at mistral.ai/products/le-chat offers free access to Small 4 with a daily usage cap. No API key required. Best for casual exploration, document summarization, and code review.
Mistral API free tier – Mistral offers a free API tier for developers with rate-limited access to mistral-small-latest, which now resolves to Small 4. Free tier requests are throttled but sufficient for prototyping and low-volume apps.
Self-hosted via Ollama or vLLM (fully free, no rate limits) – Download the weights from Hugging Face and run locally. This is the only route to zero ongoing cost and complete data privacy, but requires adequate hardware (discussed below).
OpenRouter free credits – OpenRouter occasionally provides free credits for new accounts, with Small 4 available at $0.15/M input and $0.60/M output tokens for paid usage.
For a side-by-side look at free tier limits across all major AI providers, including Mistral, check our free tier tracker which is updated weekly.
How Do You Run Mistral Small 4 Locally with Ollama?
Ollama is the simplest path to a local Mistral deployment. It wraps model downloads, quantization, and inference in a CLI with an OpenAI-compatible API. Here is the full local setup flow:
Step 1: Install Ollama – Visit ollama.com and follow the installer for macOS, Linux, or Windows. The installer includes the daemon and CLI in a single package.
Step 2: Pull the model – Run ollama pull mistral-small to download the latest quantized build. Ollama’s quantized GGUF version is substantially smaller than the raw weights and optimized for CPU+GPU hybrid inference.
Step 3: Start a chat – Run ollama run mistral-small for an interactive terminal session, or use the REST API at http://localhost:11434 for programmatic access.
Step 4 (optional): Connect to OpenAI-compatible apps – Tools like Continue.dev, Open WebUI, and LibreChat accept Ollama’s local endpoint as an OpenAI-compatible backend, letting you use Small 4 in the same interfaces you use with paid models.
For production-grade self-hosting with multi-user support and full-precision inference, vLLM is the recommended framework. The minimum infrastructure from Mistral’s own documentation is 4x NVIDIA HGX H100 GPUs, though community testers have achieved useful throughput on 2x H100 setups with the NVFP4 quantized checkpoint.
What Else Is in the Mistral 3 Family Released in Late 2025?
Mistral Small 4 is the most recent and arguably most developer-accessible release, but it builds on the Mistral 3 family announced on December 2, 2025. That announcement included:
Mistral Large 3 – The flagship open-weight model with 41B active parameters and 675B total parameters in a sparse MoE architecture. Trained on 3,000 NVIDIA H200 GPUs, it debuted at #2 in the open-source non-reasoning category on the LMArena leaderboard. Released under Apache 2.0.
Ministral 8B – An edge-optimized dense model targeting on-device and single-GPU deployments. Best performance-to-cost ratio in its parameter class at launch.
Ministral 3B – Mistral’s smallest model, designed for embedded and mobile inference. Despite its size, it outperformed many 7B-class models on instruction-following benchmarks at release.
Voxtral (March 2026) – A separate speech-generation open-source model released alongside Small 4, covering text-to-speech synthesis under the same Apache 2.0 terms. Covered separately by TechCrunch.
The full Mistral 3 family represents the most complete open-weight product lineup Mistral has released to date, covering edge deployment, production chat, reasoning, coding agents, and speech in a coherent Apache 2.0-licensed stack. For developers who have been locked into closed APIs for privacy or cost reasons, this is a meaningful alternative.
What Does the Apache 2.0 License Actually Allow Developers to Do?
Licensing matters as much as capability for serious deployments. Apache 2.0 is one of the most permissive widely-used open-source licenses. Practically, it means:
Commercial use, no royalties – You can build revenue-generating products on top of Mistral Small 4 without paying Mistral AI a cut or asking for a commercial license.
Fine-tuning for your domain – You can fine-tune the base weights on your own data and distribute or sell the resulting model. This is explicitly permitted; contrast with some other open models that restrict fine-tuned redistribution.
No network copyleft – Unlike AGPL-licensed models, running Small 4 as a hosted service does not require you to open-source your own application code.
Attribution required – You must retain the original Apache 2.0 notice and a NOTICE file if one is provided. This is a minimal obligation in practice.
This contrasts favorably with models released under more restrictive community licenses (such as Meta’s Llama license or Stability AI’s noncommercial restrictions). Apache 2.0 gives Mistral Small 4 genuine enterprise viability without the legal ambiguity that has slowed adoption of competing open models.
To explore how Mistral compares against other open-source and free-to-use models across cost, context window, and capability, visit our open-source AI hub.
🔑 Key Takeaways
Mistral Small 4 (released March 16, 2026) consolidates three separate model lines – instruct, reasoning, and coding – into a single Apache 2.0-licensed MoE model with 119B total but only 6.5B active parameters per token.
The model is genuinely free to self-host via Hugging Face and Ollama, with no royalties or commercial use restrictions under Apache 2.0 – making it one of the most permissively licensed capable AI models available today.
Performance on coding benchmarks is comparable to Claude Haiku 3.5 and Qwen 2.5 14B, while throughput improvements (40% faster completion, 3x more requests/second vs Small 3) make it compelling for production deployments.
The reasoning_effort parameter allows a single deployment to serve both fast chat and deep chain-of-thought reasoning, eliminating the need to manage multiple model endpoints for different task types.
The broader Mistral 3 family (including Mistral Large 3 at 675B total parameters under Apache 2.0) provides a full open-weight stack for developers who need to move from edge to frontier without leaving the open-source ecosystem.
Related Resources
In-depth reviews of AI tools See how the tools behind the headlines actually perform.
AI tools by profession and use case Find the right tool for what you actually do.
AI scam prevention and alerts Stay safe while exploring new AI tools.
Frequently Asked Questions
Is Mistral Small 4 completely free to use?
Mistral Small 4 is free to download and self-host under the Apache 2.0 license with no restrictions. Via Mistral’s API or services like OpenRouter, usage is metered at $0.15 per million input tokens and $0.60 per million output tokens. Mistral Le Chat offers a free web interface for casual use with a daily usage cap.
What hardware do I need to run Mistral Small 4 locally?
Mistral Small 4 has 119B total parameters, requiring significant hardware for full-precision inference. The minimum recommended setup is 4x NVIDIA HGX H100 GPUs. For more accessible hardware, quantized GGUF versions via Ollama can run on high-VRAM consumer setups (48GB+), though with reduced throughput compared to the full-precision deployment.
How does Mistral Small 4 compare to GPT-4o Mini?
Mistral Small 4 is larger and more capable than GPT-4o Mini, offering multimodal vision input, a 256k context window, and configurable reasoning modes. On coding benchmarks it performs in the same tier as Claude Haiku 3.5. The key advantage is full open-source availability under Apache 2.0, unlike GPT-4o Mini which is API-only with no self-hosting option.
What is the Apache 2.0 license and why does it matter for AI models?
Apache 2.0 is a permissive open-source license allowing free commercial use, modification, distribution, and sublicensing with no royalties. For AI models, this means you can fine-tune Mistral Small 4, build commercial products on top of it, and deploy it without paying Mistral AI or seeking permission – a significant advantage over more restrictive model licenses.
What is the reasoning_effort parameter in Mistral Small 4?
The reasoning_effort parameter toggles Small 4 between two modes per API request: set to “none” for fast, lightweight responses similar to Mistral Small 3.2, or “high” for deep step-by-step reasoning equivalent to the dedicated Magistral model. This allows one model deployment to handle both everyday chat and complex analytical tasks without switching endpoints.
Where can I download Mistral Small 4 weights?
The full model weights are hosted on Hugging Face at mistralai/Mistral-Small-4-119B-2603. Quantized builds are available through Ollama at ollama pull mistral-small. NVFP4-quantized and Eagle speculative decoding checkpoints are also available on Hugging Face for optimized inference.