Liquid AI has released two open-weight decision models aimed at edge devices. d1-3B accepts text and images. The smaller d1-omni-600M accepts text with images or text with audio, and Liquid AI describes it as an experimental research release.
The October 7 announcement reports decisions in a single forward pass, rather than answers generated token by token. Its code example asks several questions about one support ticket, including which team should handle it and how urgent it is. Another example asks the model to choose how many cats appear in an image.
Two models with different evidence behind them
d1-3B uses Liquid AI’s LFM2.5-VL-3B vision-language model as its base. d1-omni-600M uses the bidirectional LFM2.5-Encoder-350M with added vision and audio encoders. Its stated input combinations give developers separate paths for image and audio decisions.
Liquid AI reports a Decision Index 0.2.1 score of 48.57 for d1-3B, ahead of Decider 35B-A3B at 47.11. The lead chart is the company’s published comparison of index scores and model sizes.
On a separate table of seven public text benchmarks, d1-3B averages 82.9 against Decider 4B’s 81.1. d1-omni-600M averages 78.4 against Decider 2B’s 77.1. Those averages conceal different strengths. Decider 4B leads d1-3B on MASSIVE intent, BoolQ and XNLI, while d1-3B leads on the other four rows.
Liquid AI says it validated the models’ image capabilities, but publishes no vision or audio benchmark scores in this release. It also publishes no speed measurements for d1-omni-600M. Developers evaluating audio decisions therefore have less published evidence than those evaluating d1-3B’s text tasks.
Edge speed changes with the input
Liquid AI’s d1-3B timing table separates a single question, a longer text state and an image input. These are company-reported measurements.
| Device | One question | 3.4K-token state | 384px image |
|---|---|---|---|
| Apple M5 Pro | 30 ms | 640 ms | 62 ms |
| Jetson AGX Thor | 16 ms | 220 ms | 35 ms |
| Jetson AGX Orin 64 GB | 26 ms | 560 ms | 83 ms |
| Jetson Orin Nano | 50 ms | 1,640 ms | 202 ms |
The release also reports that three questions on AGX Thor take 20 ms, compared with 16 ms for one. Its examples use system_one for named questions and system_one_batch for packed requests. The setup requires transformers>=5.14 and loads the model’s own code with trust_remote_code=True.
