Perceptron: Perceptron Mk1.5

Perceptron Mk1.5 is a text model from Perceptron, listed on api14 under the model id perceptron/perceptron-mk1.5. Perceptron describes it as an embodied reasoning model for physical agents, which places it in the part of the market concerned with systems that have to interpret the physical world rather than only written language. It was released by the provider on 25 September 2026.

The model accepts text, image, video and audio as input, and it returns text. That combination matters more than it first appears: a single request can carry a written instruction alongside a camera frame, a recorded clip or a spoken command, and the model reasons across all of them together. The context window is 36,864 tokens, and the maximum output is 8,192 tokens. Beyond plain text answers, the provider describes optional structured annotations such as points, boxes, polygons and tracks, which lets the model express locations and movements in a form a program can consume directly.

Those capabilities point to a fairly specific set of uses. Robotics and embodied agent work is the obvious one, where a model has to relate language to what a sensor is showing and then emit coordinates or regions rather than a paragraph. Video and audio input also suit inspection and monitoring tasks, where the question is what is happening in a scene or a recording and where in it the relevant event occurs. Because the output can be structured, the results can feed a downstream planner, a labelling pipeline or an interface without a parsing layer in between.

You call it through api14's OpenAI-compatible API using the model id perceptron/perceptron-mk1.5, so existing OpenAI client libraries work with a change of base URL and model name. The supported request parameters are frequency_penalty, include_reasoning, max_tokens, presence_penalty, reasoning, reasoning_effort, structured_outputs, temperature, tool_choice, tools, top_k and top_p. Tool calling and structured outputs are both available, which fits the annotation use case, and the reasoning parameters let you control how much internal deliberation the model does before answering.

The main constraint to plan around is the context window of 36,864 tokens, which is modest once video or audio is involved, since media input consumes part of that budget. The output ceiling of 8,192 tokens also rules out very long generated documents in one call. If your task is pure text generation at length, or needs a much larger context, a text-only model with a bigger window will likely serve you better. Perceptron Mk1.5 is the stronger choice when the input is genuinely multimodal and the answer needs to be grounded in space, time or both.