Mistral: Mistral Medium 3.5 (batch)

Mistral Medium 3.5 is Mistral AI's dense 128B model combining instruction following, reasoning and coding in one set of weights. It receives text and images and returns text. Its maker positions it for agentic and coding tasks, while the same checkpoint can produce a direct conversational answer or use its documented reasoning mode for a more involved task.

Unified model and visual capabilities

The released card specifies 128 billion dense parameters and a 256k context window. Its vision encoder is trained to handle different image sizes and aspect ratios. Images can contribute evidence to a coding or document task: a screenshot of an interface, a diagram accompanying a requirement or a page containing information the model must interpret. The result can be a written explanation, code or structured fields rather than a generated image.

The maker describes multilingual support across dozens of languages, including English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean and Arabic. It also highlights system-prompt adherence and native function calling with JSON output. Set the audience, language and response rules separately from the material being examined so that a source document's wording remains evidence for the task.

Reasoning settings

The card documents two reasoning-effort values per request. The value none gives an instruction-style answer without reasoning mode; high enables reasoning and is recommended for complex prompts and agentic coding. This is a mode choice within the same Medium 3.5 checkpoint, rather than a different model identity.

For high reasoning effort, the maker recommends temperature 0.7 and top-p 0.95. For none, it documents temperatures between 0.0 and 0.7 depending on the task and leaves top-p unset or at 1.0. These are release recommendations. Choose the output purpose first: a concise field extraction, a code change with an explanation or a comparison that requires several justified decisions.

A practical reasoning task can supply a requirement, the implementation and the input that exposes a mismatch. Ask the model to identify the relevant path, explain the discrepancy and propose the smallest change that preserves the public interface. Request the revised code separately from the assumptions and examples. The model generates a proposal; compiling, running and applying it require the development environment connected to the agent.

Tools, documents and predictable edits

The maker's service card also lists structured outputs, predicted outputs, document questions, response prefixes, batch processing and agent conversations. For extraction, define the field names, types and treatment of missing values. JSON mode supplies a structured representation, while a custom schema defines a more specific shape in the connected interface. Ask for only the record when a program will parse it and a separate explanation when a person needs to inspect the interpretation.

Document question answering combines OCR-prepared text and structure with language-model understanding. For a comparison, identify the documents and the condition to evaluate, then request each finding with its source location. A useful output can separate shared requirements, disagreements and open questions. Direct image questions can focus on a chart label or interface element; preserving headings, units and image roles helps keep the answer tied to the supplied material.

Predicted outputs are documented for tasks where much of the desired answer is already known, such as regenerating a code file or document with limited edits. Supply the existing content and identify the specific change, preserving the parts that should remain. This is different from asking for an entirely new artifact or continuing from a known response prefix. The service features describe submission and formatting options; they do not change the model's text-and-image input or text output.

In a tool-assisted workflow, the application supplies function descriptions and parameters, the model chooses a function and prepares arguments, and the application executes it and returns the result. A calculation tool can receive values read from a supplied image, while a lookup tool can provide the missing information for the next answer. Keep the tool result associated with the operation that produced it. Agentic reasoning relies on that environment rather than on text generation alone performing external actions.

Preparing long sessions

Use the 256k context for related source code, documents and conversation history, and keep the actual goal explicit. A coding session can retain the module, its caller and the requirements while asking for one defined change. A document session can keep the relevant sections together while returning a short table or revised passage. The context capacity does not specify a separate maximum answer length; the card's 1,024-token generation example is an illustration, not that limit.

The maker notes that an older released configuration contained a long-context error and that it has been corrected, including its relevance to derived checkpoints. Preserve the corrected release configuration when using the weights for long sessions. Medium 3.5's hybrid modes and 128B architecture should not be assigned to Medium 3.1 or a separately named draft-head checkpoint merely because their names share a family.