The launch
On October 7, 2026, Liquid AI published the weights of two unusual models on Hugging Face: d1-3B and d1-omni-600M. They are "decision models": they don't generate text. You hand them a state (a support ticket, a JSON payload, a photo, or all of these) and named questions, and the model returns answers read directly off its probability distribution. The usage counter shows 0 output tokens.
Three question types
- Yes/no: returns the probability of "yes," from 0 to 1.
- Choice: picks a label from a list and returns the label, a confidence and each option's probability.
- Scale: places the state on an ordered scale of 2 to 10 levels.
Several questions about the same state are answered in one pass: per Liquid, three questions cost about 1.3 times the time of one. With no generated text, there's no broken JSON, looping or made-up format.
The two models
- d1-3B (3.12B parameters): derived from LFM2.5-VL-3B, reads text and images, 32,768-token context and 16 languages.
- d1-omni-600M (587M): an encoder that reads text, images and up to 30 seconds of audio with the same weights. It's labeled an early research release.
Numbers (Liquid's own)
- Decision Index 0.2.1: d1-3B scores 48.57, the best under 10B in the company's table and ahead of a decision model 12 times its size. Liquid computed the score itself rather than submitting it to the public leaderboard.
- Latency: about 8 ms per question on an RTX 4090 and 50 ms on the smallest Jetson Orin Nano.
- In demos, the company sorted good and defective parts from production lines at 85% to 97% accuracy with no task-specific training, and claims to match or beat frontier models on 4 of 6 tasks at 19 to 200 times lower cost. These are tests chosen and run by the vendor; there's no independent reproduction yet.
Where it runs
There's support for llama.cpp, GGUF and quantized versions, and the NVIDIA stack down to Jetson. Liquid also offers a hosted API that charges only for input tokens. The license is Liquid's own (not Apache or MIT): read the terms before commercial use.
Why it matters
Many business AI tasks don't need text: classifying tickets, approving or rejecting, spotting defects, choosing a route. Using a generative LLM for this is expensive and brittle (JSON parsing, off-format answers). A model that only decides, in milliseconds, can serve as a cheap filter before the expensive LLM call.
How to test it
- Pick a real classification task with clear labels.
- Compare accuracy, latency and cost with the LLM you use today.
- Use the returned confidence to send only uncertain cases to a human or a bigger model.
Sources
- Hugging Face, LiquidAI/d1-3B: https://huggingface.co/LiquidAI/d1-3B
- Hugging Face, LiquidAI/d1-omni-600M: https://huggingface.co/LiquidAI/d1-omni-600M
- OrcaRouter analysis (Oct 8, 2026): https://www.orcarouter.ai/blog/liquid-ai-d1-open-weights-release
Want to automate decisions and triage with AI cheaply and reliably? KGA IT offers a free consultation: kga-it.com.