Mistral Small 4 Review: What's New, Benchmarks & Pricing
Mistral released Small 4 on March 16, 2026 — a 119B MoE model that merges reasoning, vision, and agentic coding into one Apache 2.0 open-source package. Here's what changed, what the benchmarks show, and whether it belongs in your stack.
Reviewed 2026-06-11 · Source: Mistral announcement
Quick Verdict
Mistral Small 4 is the most capable truly open-source small model available as of mid-2026. The Apache 2.0 license (not the modified MIT of Medium 3.5) is the headline: you can self-host, fine-tune, and redistribute commercially without restrictions. The MoE architecture keeps inference costs low despite 119B total parameters — 6B active per token is GPU-friendly. The unified reasoning + vision + coding in one model is a genuine simplification over routing between Magistral, Pixtral, and Devstral. For developers who want frontier-adjacent capability with full model ownership, this is the best option Mistral has shipped.
What's New in Small 4
Unified reasoning, multimodal, and coding
Mistral Small 4 is the first Mistral model to consolidate Magistral (reasoning), Pixtral (multimodal), and Devstral (agentic coding) into a single set of weights. Previously you needed separate models for different task types. Small 4 handles all three without routing overhead, model switching, or latency penalties.
Configurable reasoning effort
The new reasoning_effort parameter lets you tune behavior at the API level per request. reasoning_effort="none" gives you fast, low-latency responses matching Mistral Small 3.2's conversational style. reasoning_effort="high" enables deep step-by-step reasoning equivalent to prior Magistral models. Same model, same weights — you control the quality-vs-speed tradeoff per call.
Mixture of Experts architecture at 119B
Small 4 uses a MoE architecture with 128 experts and 4 active per token. Total weight count is 119B, but only 6B parameters are active per token (8B including embedding and output layers). This means inference is far cheaper than a dense 119B model while retaining the capacity benefits of large total parameters. The same technique underpins Mixtral and GPT-4.
Native image understanding
Small 4 accepts both text and image inputs natively. Vision capabilities cover document parsing, visual analysis, and multimodal chat. This removes the need to route image tasks to a separate Pixtral model — the same API endpoint handles both modalities.
Apache 2.0 open weights
Small 4 is released under the Apache 2.0 license — a true permissive open-source license, unlike Mistral Medium 3.5's modified MIT. This means you can use it commercially, modify it, and redistribute it without additional approval. Weights are available on Hugging Face and run on vLLM, llama.cpp, SGLang, and Transformers.
NVIDIA Nemotron Coalition founding member
With Small 4, Mistral joins NVIDIA's Nemotron Coalition, signaling deep inference optimization. Small 4 has been tuned for NVIDIA infrastructure across vLLM and SGLang, and supports self-hosting on 4× NVIDIA HGX H100, 2× HGX H200, or 1× DGX B200 for the minimum viable deployment.
Benchmarks
Key performance numbers from Mistral's announcement (March 16, 2026).
| Benchmark | Small 4 | Notes |
|---|---|---|
| AA LCR (reasoning accuracy) | 0.72 | 1.6K chars output — 3.5–4x shorter than comparable Qwen models |
| LiveCodeBench (coding) | Beats GPT-OSS 120B | 20% less output than GPT-OSS 120B for equivalent score |
| End-to-end latency | 40% faster than Small 3 | Latency-optimized deployment |
| Throughput | 3× more req/s vs Small 3 | Throughput-optimized deployment |
| Context window | 256k tokens | Full long-document and agentic session support |
| Active parameters | 6B per token | MoE: 128 experts, 4 active per token (119B total) |
Pricing
Apache 2.0 license. Download from Hugging Face. Run on vLLM, llama.cpp, SGLang, or Transformers. Minimum: 4× HGX H100, 2× HGX H200, or 1× DGX B200.
Available via la Plateforme API under the Mistral Small model family. See mistral.ai/pricing for current per-token rates. Configurable reasoning_effort per request.
Available on Le Chat for conversational use. Free tier has usage limits. Pro/Team plans unlock priority access and extended context.
Who should use Mistral Small 4?
- ✓Open-source advocates and researchers — the Apache 2.0 license is genuinely permissive. Fine-tune it, modify it, ship it in commercial products without restrictions.
- ✓Regulated-industry teams — healthcare, finance, and government orgs that need on-premise LLM deployment. Self-hosting is viable on 4× HGX H100 (fewer GPUs than most 100B+ dense models).
- ✓Developers building agentic coding pipelines — Small 4 combines coding (Devstral-level), reasoning (Magistral-level), and vision (Pixtral-level) without model switching overhead.
- ✓Cost-sensitive high-throughput workloads — 6B active parameters per token means inference is dramatically cheaper than a dense 119B model. 3× throughput improvement over Small 3 compounds across large volumes.
- ✗Maximum reasoning quality — if you need the absolute best on hard math and research tasks, Claude Opus 4.8, o3, and Gemini Ultra still lead at the frontier.
- ✗Consumer assistant use cases — for everyday chat, GPT-4o and Claude Sonnet have more mature assistant features and wider ecosystem integrations.
Frequently Asked Questions
What is Mistral Small 4?
Mistral Small 4 is a 119B parameter Mixture of Experts (MoE) language model released by Mistral AI on March 16, 2026. It unifies reasoning, multimodal vision, and agentic coding into a single model under the Apache 2.0 open-source license.
How does Mistral Small 4 compare to Mistral Small 3?
Small 4 is 40% faster end-to-end and delivers 3× more throughput than Small 3 in optimized deployments. It also adds native vision support, configurable reasoning effort, and stronger coding benchmarks — capabilities that required separate models in the Small 3 era.
Is Mistral Small 4 truly open source?
Yes — it uses the Apache 2.0 license, which is a standard permissive open-source license. Unlike Mistral Medium 3.5's modified MIT license, Apache 2.0 has no additional commercial restrictions. You can use it, modify it, fine-tune it, and redistribute it freely.
What hardware does Mistral Small 4 require to self-host?
Mistral lists a minimum of 4× NVIDIA HGX H100, 2× HGX H200, or 1× DGX B200 for deployment. The recommended setup is 4× HGX H100, 4× HGX H200, or 2× DGX B200 for optimal performance. Despite having 119B total parameters, only 6B are active per token due to the MoE architecture.
Does Mistral Small 4 support images?
Yes. Small 4 accepts both text and image inputs natively. You can use it for document parsing, visual analysis, and multimodal chat through the same API endpoint.
What is the reasoning_effort parameter?
The reasoning_effort parameter controls how deeply the model reasons. Set it to 'none' for fast, lightweight responses (like Small 3.2 chat style) or 'high' for deep step-by-step reasoning (like prior Magistral models). This lets you tune the quality-vs-latency tradeoff per API request without switching models.
Try Mistral Small 4
Available as open weights on Hugging Face (Apache 2.0), via the Mistral API, and on Le Chat.
Try Mistral →Open weights on Hugging Face · Apache 2.0 license · Runs on vLLM, llama.cpp, SGLang
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.