Liquid AI's d1-2B decision model runs on a Galaxy S26 GPU, answering 4 vision questions in one pass

alexcovo_eth · x · 2026-11-09

On-device demo: LiquidAI's d1-2B decision model (built on LFM3-VL-2B), converted to LiteRT, ran on a Galaxy S26 GPU answering four questions about a photo — box counts on wheels and saddles, number of bicycles, and "Is there a car?" — in a single forward pass, timed at 1983 ms by the app.
How it works: d1-2B is a "decision model" that takes a state (string, JSON, or images) and any number of named questions, returning typed answers with probabilities without generating tokens: yes/no, choice, score, text/number/date extraction, range estimates, and object bounding boxes.
Desktop performance: the HF LiteRT repo also ships a PyTorch-free Python host; on an Apple M4 Max with Metal at float32, a single question about a support ticket takes a median 40.9 ms, and three questions in one call just 41.3 ms.
Packaging: decision graphs come in seven files for sequence lengths up to 6,144 tokens, plus variants running eight sequences in one call.

More from Infra