Mistral Small 4

An efficient open-weight model unifying instruction following, reasoning and coding.

Context256K

ModalityText · Image

AccessWeights · API

ReasoningHybrid

Where it fits.

A fit for teams seeking one smaller self-hostable model across instruct, reasoning and coding modes. Small 4 is a 119B MoE model with 6.5B active parameters, combining instruct, reasoning and coding in one efficient open model. It supports 256K context and Apache 2.0 weights.

Core capabilities

119B total parameters with 6.5B active
256K context and hybrid reasoning
Function calling, agents and built-in tools
Structured outputs, OCR, document Q&A and FIM

Best-fit workflows

Local and edge deployment
Mixed-task assistants
Cost-controlled coding

04 / OPERATIONS

From evaluation
to production.

Deployment notes

A strong candidate for controlled private inference. Start with an officially supported runtime, test both instruct and reasoning paths, and preserve an API fallback while tuning quantization, batching and tool-call parsing.

Watch before choosing

The small-model advantage depends on task decomposition and inference optimization, not parameter count alone. The small active footprint helps serving efficiency but does not remove the need to validate tool parsers, reasoning modes and quality after quantization.

Family & versions

Mistral Medium 3.5Frontier
Mistral Small 4Efficient

Specifications and positioning checked against provider documentation.