Liquid AI has open-sourced the weights of two small decision-making models: the 3B parameter version delivers answers in 8 milliseconds.

Beating AI Express: After releasing its cloud-based decision-making model d1 earlier, Liquid AI has opened the weights of two smaller decision models: d1-3B and the experimental d1-omni-600M. The two models have approximately 3.1 billion and 600 million parameters respectively, enabling developers to download, fine-tune, and deploy them on local devices. The cloud-based d1, which previously scored 58.9 in comprehensive evaluations, remains weight-locked. The d1-3B supports text and image inputs, capable of judging content compliance, selecting tools for agents, or making judgments based on visuals. In Decision Index v0.2.1 benchmarks, the official self-test scored 48.57 points, outperforming Decider 35B-A3B (47.11 points) — a model with roughly 12 times its total parameters. On an RTX 4090, it takes just 8 milliseconds to answer a single short question; even on NVIDIA’s small AI computer for robots, inference takes only 50 milliseconds. Processing long texts or images, however, will result in longer latency. The experimental d1-omni-600M has around 600 million parameters and supports direct intent recognition of voice commands, eliminating the need for prior text conversion. It is better suited for small devices with limited resources but is still in early development. In the Decision Index text decision benchmark, it scored just 15.95 points. The official has not yet released the model’s inference speed or independent visual and speech evaluation results. Weights for both models are now available on Hugging Face. However, they are governed by the LFM Open License v1.0, which carries commercial use restrictions. Enterprises with annual revenue of $10 million or more must obtain separate authorization for commercial applications.