Mistral Small 4

Mistral Small 4 represents a deliberate consolidation effort, merging the formerly separate Mistral Instruct, Magistral reasoning, and Devstral coding lines into a single hybrid model rather than asking developers to route between specialists. The architectural backbone is a 119B-parameter mixture-of-experts design with 128 experts and 4 active per token, yielding roughly 6.5B activated parameters per forward pass. A 262k token context window, multimodal input handling for text and images, and configurable reasoning effort per request make it a flexible general-purpose system that can switch between quick conversational replies and more deliberate analytical responses without changing weights or endpoints.

For practitioners, the practical appeal lies in deployment efficiency and openness. Mistral reports a 40% reduction in end-to-end completion latency and roughly 3x higher requests-per-second throughput compared to Mistral Small 3 in optimized serving setups, and the team ships companion artifacts — a trained Eagle speculative-decoding head and an NVFP4 4-bit float checkpoint — that let operators trade memory for speed without retraining. The model fits well as a single backbone for production assistants that need chat, image understanding, tool-augmented reasoning, and lightweight code generation on the same API call, particularly for teams that want an open-weights alternative to comparable dense frontier models and value the ability to self-host quantized variants for cost control.

OpenRoutermistralai/mistral-small-2603mistral-small

Open provider page Provider docs

Quick Info

Provider: OpenRouter
Model key: mistralai/mistral-small-2603
Release date: Mar 16, 2026
Last updated: Mar 16, 2026
Knowledge cutoff: 2025-06
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.15
Output token cost: $0.60

Limits

Output tokens: 209,715 tokens
Context window: 262,144 tokens

Transparent token rates

Compare Mistral Small 4 pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Mistral Small 4

OpenRouter

CoverageRelease Notes

Mistral AI Releases Mistral Small 4: A 119B-Parameter MoE Model that Unifies Instruct, Reasoning, and Multimodal Workloads

OpenRouter

Coverage

Mistral AI has released Mistral Small 4, combining fast text responses, logical reasoning, and image processing in one model.

OpenRouter

CoverageRelease Notes

NVIDIA's NeMo AutoModel release log records that Mistral Small 4 was added on 2026-03-16 under the Hugging Face identifier `mistralai/Mistral-Small-4-119B-2603`, listed as a Vision-Language Model (VLM) and accompanied by a recipe example (`mistral4\ medpix.yaml`) that developers can run. The log entry is a reverse-chro For practitioners, the NeMo entry means Mistral Small 4 has a first-party reference recipe within an NVIDIA-maintained training stack, enabling reproducible fine-tuning and evaluation workflows rather than ad hoc setups. Combined with the explicit HF model ID naming the exact 2603 variant, the documentation confirms Mi

OpenRouter

Coverage

Pi's developer integration page for Mistral Small 4 confirms it is served via OpenRouter's OpenAI-compatible completions endpoint at https://openrouter.ai/api/v1, with a 262,144-token context window, 4,096 max output tokens, and reasoning support enabled. Modalities are listed as text and image input, matching the prim The page enumerates effective compatibility flags that matter for tool-use integrations: supportsStore is true, supportsDeveloperRole is false, supportsReasoningEffort is true, supportsUsageInStreaming is true, maxTokensField maps to max completion tokens, requiresToolResultName is false, requiresAssistantAfterToolResu

OpenRouter

Official sourceBenchmark

Mistral Small 4 (mistralai/mistral-small-2603) is available on OpenRouter as a unified release that combines reasoning capabilities from Magistral, multimodal understanding from Pixtral, and agentic coding from Devstral into a single model. According to the OpenRouter listing, it was released on March 16, 2026, support OpenRouter currently routes the model through four upstream providers: Mistral (ZDR), Mistral, Mistral (EU), and Venice, each with different latency, throughput, and uptime characteristics. Provider-level performance figures include Mistral ZDR at 0.78s latency and 72 tok/s throughput with 99.47% uptime, while Venice s