Liquid AI's d1-3B tops Decision Index 0.2.1 at 48.57

Liquid AI built these two models on its Liquid Foundation Models, but with a twist: unlike generative models that produce tokens, decision models answer in a single forward pass. That design choice is what makes the edge-speed numbers possible. d1-3B is trained from LFM2.5-VL-3B, a decoder-only vision-language model, while the experimental d1-omni-600M is trained from LFM2.5-Encoder-350M, a bidirectional encoder that adds vision and audio encoders to handle all three modalities. The two backbones point at different use cases: d1-3B for text-and-image decisions where quality matters, and d1-omni-600M where footprint is the constraint.

The benchmark results show d1-3B leading on mean score across seven public datasets covering reading comprehension, toxicity detection, intent classification, medical QA, and cross-lingual understanding. The company did not report vision or audio benchmarks, noting that the Decision Index v0.3 includes only a private vision split and that audio decision benchmarks are currently an open problem. That gap matters for anyone hoping to compare the multimodal claims directly against rivals.

The speed validation, done with NVIDIA across GeForce RTX 4090, Jetson AGX Thor, Jetson AGX Orin 64 GB, and Jetson Orin Nano, shows d1-3B answering a single question in under 50 ms on every measured device. Whether that translates into real-world product wins will likely depend on how the models perform on each team's specific decision tasks, and on when d1-omni-600M emerges from its early research release.