Liquid AI releases d1-3B and d1-omni-600M for single-pass edge decisions

Liquid AI released two downloadable “decision models” on October 7: d1-3B and the experimental d1-omni-600M. Instead of generating text token by token, they return typed answers—yes/no probabilities, named choices or ordered scores—in a single forward pass. That design targets routing, moderation, triage, reranking, guardrails and visual inspection where a structured decision matters more than prose.

The 3.12-billion-parameter d1-3B accepts text, JSON and images, supports a 32,768-token context, and is available on Hugging Face with Transformers and llama.cpp support. The smaller d1-omni-600M accepts text with either images or audio, but Liquid calls it an early research checkpoint and does not publish latency measurements for it.

Liquid reports d1-3B scoring 48.57 on the public split of Decision Index 0.2.1, ahead of the compared models under 10B and close to the much larger Decider 35B-A3B at 47.11. The company measured 8 ms for one warm question on an RTX 4090 and 50 ms on a Jetson Orin Nano. Those latency results cover one request at a time and specific input shapes; the RTX result also uses CUDA graphs, and Liquid says the first call with a new shape pays compilation or kernel-selection overhead.

The benchmark evidence is promising but limited. Liquid ran its own models with the official scorer rather than submitting them to the public leaderboard. A private vision split was not disclosed, and the company says dedicated audio-decision benchmarks remain an open problem.

Both checkpoints use the custom LFM Open License v1.0, not Apache or MIT. It permits modification and commercial use below $10 million in annual revenue; organizations at or above that threshold need a separate commercial license.

Sources: Liquid AI announcement · d1-3B model card · d1-omni-600M model card · LFM license · Unsplash image license