Liquid AI's d1 Decision Models Went Open: Triage Support Tickets on a CPU With the 600M One (and Where It Fools You)

Liquid AI recently made its d1 decision models open source, allowing users to run text, vision, and audio-based decision models on a single CPU. The company released two models: d1-3B (3.12B parameters, built on LFM2.5-VL-3B) and d1-omni-600M (350M bidirectional encoder plus vision and audio encoders). The d1 models return typed, calibrated answers in one forward pass, without generating any output tokens. This is useful for tasks such as ticket routing, intent classification, and moderation checks.

Liquid AI claims that the d1-3B model scores 48.57 on Decision Index v0.2.1, which is the best under 10B and on par with a 35B-A3B decision model. The d1-omni-600M model scores 15.95 on the same index and is described as an early research release, with no speed numbers provided. Both models support day-one transformers via trust_remote_code and have llama.cpp support. The 3B model is the one Liquid AI is pushing.

Liquid AI tested the 600M model by running 12 short support tickets through the system and measuring latency, accuracy, and batch performance. The results showed that the model could handle mixed tickets and flag high-risk cases with confidence scores. However, the model also incorrectly classified some non-ticket inputs as billing or technical issues. To address this, Liquid AI recommends adding a catch-all option and a separate noul gate to ensure more accurate classification.