Mistral's 2026-03 unified small model (119B MoE, 6B active). Combines Magistral (reasoning), Pixtral (multimodal), and Devstral (agentic coding) capabilities into a single model.
Mistral Small 4 is a 119-billion-parameter Mixture-of-Experts (MoE) model released on March 16, 2026, under Apache 2.0. With only 6.5 billion active parameters per token (out of 128 experts with 4 active per token), it unifies four previously separate Mistral model families — Magistral (reasoning), Pixtral (multimodal), Devstral (agentic coding), and Mistral Small (instruct) — into a single deployable model. This is Mistral's first model to consolidate their entire non-flagship stack into one architecture, dramatically simplifying routing logic for production deployments.
For a TheRouter operator, Mistral Small 4 is the cost-effective 'Swiss Army knife' — it handles vision, tool calling, reasoning, and coding from a single endpoint at $0.20/M input tokens (via TheRouter). The configurable reasoning_effort parameter (none/high) lets you dial compute per-request, making it equally suitable for fast chat and deep step-by-step reasoning without maintaining separate model deployments. It supports 256K token context (upstream) and is self-hostable on modest multi-GPU setups (4× H100 or 2× H200 minimum).
| Type | Rate |
|---|---|
| Input | $0.162 / 1M tokens |
| Output | $0.648 / 1M tokens |
| Release date | 2026-03-16mistral.ai ↗ | verified |
| Architecture | 119B total / 6.5B active — MoE (128 experts, 4 active per token); Transformer with grouped-query attention (GQA)docs.mistral.ai ↗ | verified |
| Upstream context length | 262,144 tokens (256K); TheRouter serves 131,072 per standard-models.yamlhuggingface.co ↗ | verified |
| Training cutoff | Not publicly disclosed | unknown |
| License | Apache 2.0 (commercial use permitted, no usage caps)huggingface.co ↗ | verified |
| Supported inference backends | vLLM, SGLang, llama.cpp, Transformers, NVIDIA NIMhuggingface.co ↗ | verified |
| Minimum self-host hardware | 4× NVIDIA HGX H100, 2× HGX H200, or 1× DGX B200mistral.ai ↗ | to verify |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
AA LCR Academic Agent LiveCodeBench Reasoning. Score achieved with ~1.6K characters output — notably concise vs comparable Qwen models needing 5.8-6.1K chars for similar scores | 0.72 | mistral.ai ↗ | |
LiveCodeBench Mistral-reported; beats GPT-OSS 120B while producing 20% less output | Outperforms GPT-OSS 120B | mistral.ai ↗ | |
AIME 2025 American Invitational Mathematics Examination 2025 — matches GPT-OSS 120B with shorter output | Matches GPT-OSS 120B | mistral.ai ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "mistral/mistral-small-4",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Basic chat completion to test the endpoint or run conversation tasks.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "mistral/mistral-small-4",
"messages": [
{"role": "user", "content": "Explain the difference between MoE and dense transformer models"}
]
}'Mistral AI released Mistral Small 4 (mistral-small-2603), a 119B MoE model unifying Magistral reasoning, Pixtral multimodal, Devstral coding, and Mistral Small instruct capabilities into a single open-weight (Apache 2.0) deployment. Key features include configurable reasoning_effort, 256K context, and self-hostability on 4× H100 GPUs.
Mistral Small 4 is a fundamentally different architecture: it is a 119B MoE with 6.5B active parameters, not a dense 24B model. It unifies four previously separate Mistral model families (Magistral, Pixtral, Devstral, Mistral Small) into one. It adds vision support, configurable reasoning_effort, and a 256K context window. Mistral claims 40% lower latency and 3× higher throughput vs Small 3.
Yes — TheRouter provides full OpenAI-compatible endpoints at api.therouter.ai/v1. Just set the baseURL and your TheRouter API key, and specify model: 'mistral/mistral-small-4'. The reasoning_effort parameter is passed as a top-level request field — some SDKs may need @ts-expect-error if they don't forward unknown keys.
The model weights are released under Apache 2.0 license — this is open-weight, meaning you can deploy, fine-tune, and redistribute the weights without restrictions. It requires at least 4× NVIDIA H100, 2× HGX H200, or 1× DGX B200 for self-hosting. Supported serving frameworks include vLLM, SGLang, llama.cpp, and Transformers.
Set reasoning_effort to 'none' for fast, direct responses (comparable to Mistral Small 3.2 speed). Set it to 'high' to activate step-by-step reasoning equivalent to Magistral-level depth. This eliminates the need to maintain separate fast and reasoning model deployments — one model handles both modes. The parameter is passed as a top-level field in the chat completion request body.
Mistral reports that Small 4 achieves AA LCR score of 0.72 (with notably concise output ~1.6K chars vs 5.8-6.1K for comparable Qwen models). It outperforms GPT-OSS 120B on LiveCodeBench while producing 20% less output, and matches GPT-OSS 120B on AIME 2025. More detailed third-party benchmark data is expected as independent evaluators publish their results.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | mistral.ai ↗ | 2026-05-27 | verified |
| Architecture | docs.mistral.ai ↗ | 2026-05-27 | verified |
| Upstream context length | huggingface.co ↗ | 2026-05-27 | verified |
| Training cutoff | — | — | unknown |
| License | huggingface.co ↗ | 2026-05-27 | verified |
| Supported inference backends | huggingface.co ↗ | 2026-05-27 | verified |
| Minimum self-host hardware | mistral.ai ↗ | 2026-05-27 | to verify |
| AA LCR | mistral.ai ↗ | 2026-05-27 | to verify |
| LiveCodeBench | mistral.ai ↗ | 2026-05-27 | to verify |
| AIME 2025 | mistral.ai ↗ | 2026-05-27 | to verify |
| Introducing Mistral Small 4 | mistral.ai/news ↗ | 2026-05-27 | verified |
| What makes Mistral Small 4 different from Mistral Small 3? | mistral.ai ↗ | 2026-05-27 | to verify |
| Can I use Mistral Small 4 via the standard OpenAI Python/TypeScript SDK? | docs.therouter.ai ↗ | 2026-05-27 | to verify |
| Is Mistral Small 4 open-source? Can I run it on my own hardware? | huggingface.co ↗ | 2026-05-27 | to verify |
| How does the reasoning_effort parameter work? | docs.mistral.ai ↗ | 2026-05-27 | to verify |
| What are the benchmark scores for Mistral Small 4? | mistral.ai ↗ | 2026-05-27 | to verify |