Mistral: Mistral Medium 3.5 (batch)

Mistral Medium 3.5 (batch) is a text generation model from Mistral AI, identified on api14 as mistralai/mistral-medium-3-5:batch. It is a dense 128B instruction-following model, which means all of its parameters are active on every request rather than being routed through a mixture-of-experts setup. The batch variant is the version exposed through this endpoint, and the model sits in the text category on api14.

The model accepts text, image, and file inputs, and it returns text only. Its context window is 262,144 tokens, and the maximum output length is 209,715 tokens, so it can take in very large inputs and produce long responses in a single call. Because images and files are accepted alongside text, you can send mixed prompts that combine written instructions with visual or document content, and the model will reason over all of it before generating its answer.

Those capabilities point to a few practical uses. The large context window suits work over long documents, codebases, or multi-file inputs where you would otherwise need to chunk and stitch results. Image input makes it a fit for tasks that mix screenshots, diagrams, or scanned pages with text instructions. Mistral describes the model as designed for agentic workflows, coding, and complex tasks, so tool calling and structured generation are reasonable starting points. The batch variant is a natural choice for offline or throughput-oriented jobs where you submit many requests and collect the results, rather than for interactive chat where latency matters most.

You call the model through api14's OpenAI-compatible API using the model id mistralai/mistral-medium-3-5:batch. The endpoint supports frequency_penalty, include_reasoning, max_tokens, presence_penalty, reasoning, reasoning_effort, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, and top_p. In practice that means you can control sampling with temperature, top_p, and the two penalty parameters, cap output with max_tokens, fix randomness with seed, and define stop sequences. For structured work, response_format and structured_outputs let you constrain the shape of the reply, while tools and tool_choice support function calling, and the reasoning parameters let you tune how much internal reasoning the model exposes or performs.

One thing to weigh is the batch designation. If your application needs low-latency, turn-by-turn responses, a non-batch model is likely the better fit, and the same applies if you only need text input and output, since the image and file modalities would go unused. The 262,144-token context and 209,715-token output ceiling are generous, but they are still finite, so extremely large corpora may need to be split across calls. Within those bounds, the model covers long-context text work, multimodal prompts, coding, and tool-driven agent flows from a single endpoint.