Mistral: Mistral Small 4 (batch)

Mistral Small 4 combines general instruction following, reasoning and coding in a single multimodal model. The maker describes it as unifying the Instruct, Magistral reasoning and Devstral coding families. It accepts text and images and returns text, with function calls and JSON output available for tool-assisted applications.

Architecture and key features

The release has 119 billion total parameters and activates 6.5 billion per token. Its mixture-of-experts architecture contains 128 experts, with 4 active, and the maker documents a 256k context window. The active parameter count and number of experts describe how the model is organized; they are distinct from the amount of text or image content included in one task.

The model provides instruction and reasoning functionality in the same checkpoint. The maker also identifies vision, system-prompt adherence and multilingual support across dozens of languages, including English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean and Arabic. A text instruction can therefore ask about an image or combine a written requirement with visible evidence, with the answer returned as text.

A visual task can provide a document page and ask for the values relevant to a named condition. Keep the labels and units readable and state whether the result should preserve the original wording, summarize it or compare it with another source. An interface task can pair a screenshot with the user's goal and request the relevant visible control and the evidence supporting a proposed next step.

Choosing instruction or reasoning mode

The released card documents reasoning effort none for an immediate instruction-style answer and high for reasoning, with high recommended for complex prompts. The choice is made per request within the same Small 4 model. An ordinary classification or short rewrite can have a different answer contract from a mathematical, visual or coding problem that needs a developed explanation.

The maker recommends temperature 0.7 for high reasoning effort. For none, it specifies temperatures between 0.0 and 0.7 according to the task. These settings describe the documented modes and do not turn reasoning effort into a guarantee of a correct result. Define the assumptions, output form and relevant source material before choosing how much reasoning the task requires.

For a worked problem, provide the full condition and ask for the important justified steps followed by the final expression. For a code task, provide the requirement, the affected function and a concrete input; request the proposed implementation separately from the explanation and examples. The maker's usage demonstrations include instruction following, a calculation tool applied to an image and visual reasoning, illustrating how text and visual evidence can enter these different tasks.

Coding, documents and tools

The model is positioned for codebase exploration and software-engineering agents, general assistants, document understanding and research. A coding-agent request can name the module to change, the interface to preserve and the behavior that should result. Supply the relevant source and failure evidence. The model can propose the next operation or a code change, while the application provides repository access, execution and feedback.

For document extraction, identify the required fields and how to represent missing information. A useful record can separate the document identifier, source statement, interpreted requirement and unresolved question. The maker's service card lists custom structured output and document question answering; the latter uses document processing and OCR to prepare text and structure before language-model analysis. A direct image question and an OCR-assisted document workflow should retain their own input preparation.

Function calling requires clearly described tools. The model selects a function and creates arguments; the application executes the operation and returns its result for the subsequent answer. The same logic applies whether the task is a lookup, a calculation or a coding operation. JSON formatting organizes an answer, while a tool call asks for an external action; the two should not be treated as interchangeable results.

The service card also lists predicted outputs, response prefixes, batching and agent conversations. Predicted outputs are useful when a document or code file must be regenerated with only a limited change and the rest of the content is known. Specify the edit and the material to preserve. Prefix continuation uses a known beginning of the response, while batching concerns submission of work rather than an additional reasoning mode.

Context and edition boundaries

Plan the written task, visual material, conversation and desired answer around the documented 256k context. The released usage example configures a 262,144-token context, while its 1,024-token generation example is not a separate maximum output capacity. The sources read do not provide an independent hard answer-length limit to substitute for the chosen task length.

Small 4 has its own 119B mixture-of-experts architecture and hybrid modes. It should not inherit the dense 128B architecture of Medium 3.5 or the separate Instruct and Reasoning editions of Ministral 3. The maker also releases distinct quantized and speculative-decoding artifacts; those are separate deployment artifacts rather than different content modalities. Preserve the named Small 4 release and the mode requested for the actual explanation, structured record or code result.