Skip to main content

​
2026-09-25 — Perceptron Mk1.5

Use the model ID perceptron-mk1.5 with the chat completions API.
  • Audio input: transcribe, summarize, and ask questions using input_audio, audio_url, or audio_file_id with WAV, MP3, or FLAC. The current per-item limit is 16,384 audio tokens, approximately 21.8 minutes; audio also occupies the shared 36,864-token context. Usage is reported in usage.prompt_tokens_details.audio_tokens. See Audio.
  • Video soundtracks: set vision_config.enable_audio_in_video: true to analyze audio alongside video frames. This is off by default. See video soundtracks.
  • Reasoning effort: use none, minimal, low, medium, or high on chat completions and Multilook. vision_config.enable_thinking is deprecated. See Reasoning.
  • Function calling: declare functions, receive tool calls, and return results for the next turn. See tool calling.
  • Video tracking: request <track> output with timestamped spatial observations. See video tracking.
  • Multiple assets: the optional asset_idx attribute associates annotations with an input across images, videos, and conversation history. See multiple assets for handling both present and omitted selectors.
  • Structured output guidance: distinguish native annotations, constrained final responses, and function arguments. See structured outputs.
  • Hosted MCP: question and caption accept modality: "audio" and opt-in video soundtracks; all four inference tools accept reasoning_effort.
  • MCP model default: question, caption, ocr, and detect now default to perceptron-mk1.5 when model is omitted. Pass model: "perceptron-mk1" to retain the previous model. See the MCP guide for hosted access and npm bridge compatibility.
The model card lists supported inputs, context limits, and pricing.