Liquid AI Opens Weights for Two Small Decision Models: 3B Version Delivers Answers in 8 Milliseconds

Beating AI News Flash: Following the earlier release of its cloud-based decision model d1, Liquid AI has now opened the weights for two small decision models: d1-3B and an experimental version, d1-omni-600M. The two models have approximately 3.1 billion and 600 million parameters, respectively, and developers can download, fine-tune, and deploy them on local devices. The cloud-based d1, which previously scored 58.9 overall, has not yet had its weights opened.

d1-3B supports text and image inputs and can determine whether content violates rules, select tools for an Agent, or make judgments based on visuals. In the Decision Index v0.2.1 evaluation, official self-testing yielded a score of 48.57, surpassing Decider 35B-A3B (47.11 points), which has about 12 times its total parameters. On an RTX 4090, it takes only 8 milliseconds to answer a single short question. Even on NVIDIA's small AI computer for robots, it takes only 50 milliseconds. However, processing long texts or images takes longer.

The experimental d1-omni-600M has only about 600 million parameters and also supports directly judging the intent of voice commands without first converting them to text. It is better suited for small devices with limited resources, but it is still in an early stage. In the Decision Index text decision evaluation, it scored only 15.95. Officials have not yet released the model's inference speed or independent visual and speech evaluation results.

The weights for both models have been uploaded to Hugging Face. However, the LFM Open License v1.0 they use has commercial restrictions. Enterprises with annual revenue reaching $10 million need to obtain separate authorization for commercial use.