Rather than generating text token by token, these ‘decision models’ return a short, structured response in a single inference, for real-time embedded applications.
Liquid AI has uploaded two open-weight models of a particular kind to Hugging Face. Rather than producing a word-by-word response, its "decision models" render a brief, structured response (125 characters or less) in a single pass, making them suitable for routing, flow control, or immediate reactions rather than conversation.
Two sizes, three modalities
d1-3B, derived from the LFM2.5-VL-3B vision-language model, accepts text and images. The smallest, d1-omni-600M, is based on a bidirectional encoder of 350 million parameters to which the publisher has grafted visual and audio encoders; it is presented as experimental. On Liquid AI's decision index, d1-3B obtains 48.57 points, a score that the company considers comparable to that of a competing model twelve times larger.
Latency is the central argument: about 8 ms per question on an RTX 4090 card, 26 ms on a Jetson Orin module and 50 ms on a Jetson Orin Nano. Liquid AI cites gesture-driven games, live content moderation or robot browsing as use cases.
Read the full article on the original source
Go to the source →