Mistral: Mistral Small 3.2 24B

Mistral Small 3.2 24B is an instruction-tuned language model from Mistral, released by the provider on 2025-06-20 under the model id mistralai/mistral-small-3.2-24b-instruct. It is a 24B parameter model, and this 3.2 release is an update to the 3.1 version, optimized for instruction following, repetition reduction, and improved function calling. It sits in the text category, and its knowledge cutoff is 2023-10-31.

The model accepts both text and images as input and produces text as output, so it can handle multimodal prompts that combine written instructions with visual content. It has a context window of 256,000 tokens, which allows long documents, extended conversations, or large image and text payloads to be processed in a single request. The maximum output length is 16,384 tokens. Supported request parameters include temperature, top_p, top_k, min_p, frequency_penalty, presence_penalty, repetition_penalty, max_tokens, stop, seed, logit_bias, logprobs, top_logprobs, response_format, structured_outputs, tools, and tool_choice.

Those capabilities point to a few practical uses. The long context and image input suit document analysis, where a developer sends a scanned page or screenshot alongside questions about its contents. The improved function calling and tool support make it a reasonable choice for agent-style workflows that need to call external functions or APIs. Structured outputs and response_format help when the application needs JSON or another predictable shape rather than free text. Instruction following and repetition reduction matter for chat assistants, summarization, and extraction tasks where consistent phrasing is important. The seed parameter supports reproducible sampling for testing.

You call the model through api14's OpenAI-compatible API using the model id mistralai/mistral-small-3.2-24b-instruct. Because the interface follows the OpenAI chat completions format, you send a request with the model field set to that id and a messages array containing your text and image content. The request parameters listed above are supported, so you can set sampling controls such as temperature, top_p, top_k, and min_p, apply penalties, cap output with max_tokens, constrain generation with stop, response_format, or structured_outputs, and attach tools with tool_choice for function calling. Log probability options (logprobs and top_logprobs) and logit_bias are also available.

Two limits are worth noting when deciding. The knowledge cutoff is 2023-10-31, so questions about events after that date rely on information you supply in the prompt rather than the model's own knowledge. Output is capped at 16,384 tokens, which may be tight for tasks that need very long generated documents. If your workload is text only, or if you need a model with a more recent cutoff, another model may fit better.