Liquid AI has published the weights for two d1 models on Hugging Face, so anyone can now download them and run them locally. The release turns d1 from a product claim into something outsiders can test, starting with d1-3B and its smaller sibling d1-omni-600M.
A decision model, as Liquid AI describes it, does not write text. It reads an input (text, an image, or audio) and returns its answer in one computation, without generating tokens one at a time. The company’s text tests cover medical question answering, cross-lingual understanding, intent classification, reading comprehension, and toxicity detection. Those are sorting and judging jobs, not drafting jobs. Liquid AI does not offer a tighter definition in the post.
Our earlier coverage of d1 rested on the company’s comparisons against hosted frontier models. This release is a different event: the weights themselves. Liquid AI’s blog post, published 7 October, calls the models open-weight and says users can “download, fine-tune, and deploy without restrictions.” The post does not name a licence, so legal teams should read the terms on the Hugging Face pages before assuming anything beyond that sentence.
The numbers are Liquid AI’s own. d1-3B scores 48.57 on the public split of the company’s Decision Index, version v0.2.1, which the company says is ahead of every model under 10 billion parameters and level with Decider 35B-A3B, a model twelve times its size. On seven public text benchmarks, d1-3B averages 82.9, against 81.1 for the next-best Decider 4B. No independent evaluation accompanies any of this, and the Decision Index is a benchmark the company reports on itself.
The speed claims are the more checkable part. Liquid AI says d1-3B answers one question in 8 milliseconds on an NVIDIA RTX 4090, 16 on a Jetson AGX Thor, and 26 on a Jetson AGX Orin. It reports 30 milliseconds on an Apple M5 Pro and 50 on the smallest board it tested, the Jetson Orin Nano. Give that board a long input, a 3,400-token state, and the same table shows 1,640 milliseconds. Latency depends heavily on how much you feed it.
The 600M model is the one to treat carefully. Liquid AI calls d1-omni-600M its first experimental checkpoint and an early research release. It accepts text with an image, or text with audio. It scores 15.95 on the Decision Index, a third of d1-3B’s figure, and the company published no latency numbers for it. On the text benchmarks it averages 78.4, which beats the 2B Decider’s 77.1 and tops the table on toxicity detection and paraphrase identification.
That mix is where the open release matters. Plenty of teams cannot send customer messages, clinical notes, or camera feeds to an outside API, and for them a classifier that fits on a Jetson board is the difference between using AI on that data and not using it at all. A 600M-parameter model that handles audio and images and sits entirely inside a company’s own network is a real option for that group, even in its rough first form. Because the weights are open, a team can also fine-tune it on its own labelled examples, which is usually what a narrow classification job needs.
The practical step is small: pull d1-3B, run it on a few hundred of your own labelled examples, and compare its accuracy and latency with whatever API you pay for today. If it holds up on your data, the per-call bill for that workload drops to the cost of the hardware.
Liquid AI, “Open d1: Edge decision models for text, vision, and audio,” Liquid AI blog, 7 October 2026.
