Perceptron has shipped Mk1.5, a compact multimodal reasoning model tuned for physical agents — drones, quadrupeds and smart glasses — that the ex-Meta FAIR team claims outperforms egocentric AI incumbents at 25x lower cost. Released September 25, 2026 on OpenRouter at $0.15/$1.50 per million input/output tokens, Mk1.5 also lands on AWS Marketplace, a distribution push that follows the company's $21M Bessemer round and Isaac 0.5 launch in August.
What Is New Under The Hood
Mk1.5 processes text, image, video and audio in a 36,864-token context, delivers up to 8,192 tokens of output, and returns structured annotations — points, bounding boxes, polygons, motion tracks and video clip references. On the P50, Perceptron measures 0.50 s latency and 106 tokens per second throughput on its own hosting. Compared with Mk1, the company reports a 2x–5x latency reduction and a new state-of-the-art in egocentric understanding.
Audio As First-Class Sensor
The most unusual addition is elevating audio to a primary modality. Mk1.5 can, for example, infer an object's proximity from the Doppler effect and use ambient sound to disambiguate scenes when vision is degraded. It also supports zero-shot control of quadrupeds and drones without platform-specific fine-tuning, and can spawn parallel sub-agents that decompose complex on-device tasks.
Why It Matters
The pitch — frontier egocentric reasoning at open-weight-competitive pricing — targets a market where Google's Gemini Live and OpenAI's realtime multimodal models still dominate but at costs that make continuous on-device inference commercially difficult. Perceptron argues the combination of visual grounding, tool calling and structured outputs lets fleet operators put an actual reasoning loop into industrial drones and inspection robots instead of streaming everything back to a cloud.
Where It Fits In
Mk1.5 slots below NVIDIA Isaac ROS 5.0 in the stack — Perceptron provides the perception-and-reasoning brain; NVIDIA and Google supply the runtime and hardware. The rollout also comes as Black Forest Labs' FLUX 3-Action and Skild AI's S1 vie for the same physical-AI foundation-model tier that Perceptron is pushing into.
Reporting based on coverage from HuggingNews, OpenRouter and Perceptron.