The question before the model
Does a device that only needs to decide something need to write a sentence first? Liquid AI's answer, published October 7, 2026, is two open-weight models in its d1 decision family: d1-3B and an experimental d1-omni-600M. Liquid AI says that unlike its generative models, decision models do not produce tokens. They answer in a single forward pass. The company's own code example shows the shape of the work: given a customer message, the model answers a yes-or-no question about whether a refund is requested, a choice question about which team should handle it, and a score question about urgency on a three-step scale, all in one call.
Source: Multimodal open d1 decision models for the edge · 2026-10-07 · Primary source context · attributed account
Two backbones, two sets of inputs
The two models come from different places. Liquid AI says d1-3B is trained from LFM2.5-VL-3B, a decoder-only vision-language model, and takes text and images. The smaller d1-omni-600M is trained from LFM2.5-Encoder-350M, a bidirectional encoder, with vision and audio encoders added, and takes either text and image or text and audio. Liquid AI describes d1-omni-600M as an early research release still under development, which is why it reports no speed numbers for it. The company also reports no vision or audio benchmarks for either model, saying the vision split of the Decision Index v0.3 is private and that audio decision benchmarks are currently an open problem.
Source: Multimodal open d1 decision models for the edge · 2026-10-07 · Primary source context · attributed account
What Liquid AI reports, and on what
All numbers here are Liquid AI's, on its chosen datasets and hardware. Across seven public datasets covering reading comprehension, toxicity detection, intent classification, medical question answering and cross-lingual understanding, the company reports a mean of 82.9 for d1-3B and 78.4 for d1-omni-600M. On the Decision Index 0.2.1, Liquid AI reports d1-3B at 48.57, which it says puts the model ahead of every 4B and 9B model and of Decider 35B-A3B at 47.11. For speed, measured with NVIDIA, d1-3B answers one question in 16 milliseconds on a Jetson AGX Thor, 26 on a Jetson AGX Orin and 50 on a Jetson Orin Nano. Three questions together cost about 1.3 times one, the Thor moving from 16 to 20 milliseconds.
Source: Multimodal open d1 decision models for the edge · 2026-10-07 · Primary source context · attributed account
What a decision model is not
Silver Owl analysis: a model that returns a yes, a choice or a point on a scale cannot explain itself in prose, so the person reading the output has to supply the context. That narrow shape is the attraction for a small team sorting messages or images into bins, and also the limit. One hypothesis we would want tested before relying on it is whether response times stay steady under a real queue of mixed inputs, since Liquid AI's published figures come from its own measurements on specific boards. The open weights on Hugging Face make that test possible today.
Source: Multimodal open d1 decision models for the edge · 2026-10-07 · Primary source context · attributed account