Liquid AI Opens Small Models That Decide Without Talking

Most of the attention in AI goes to models that talk: the bigger the vocabulary of plausible sentences, the better. Liquid AI has just released two open-weight models built on the opposite idea. They do not generate text at all. They decide.

On Tuesday the company put d1-3B and an experimental d1-omni-600M on Hugging Face as open-weight downloads, described in a post on the Liquid AI blog. Both belong to what Liquid calls its d1 family of decision models. Rather than writing out a response token by token, each model answers in a single forward pass and returns a structured result: a yes or no, a choice from a list, or a score against a rubric. The three can be combined in one call.

Why that matters

Generating text is expensive. A model that has to spell out its reasoning word by word burns time and compute on every request, which is a poor fit for the dull, high-volume decisions that fill real systems. Is this email spam? Does this image contain a defect? Which of these five categories applies? For those jobs you do not want an essay. You want an answer, immediately.

Liquid says d1-3B can produce one in about 8 milliseconds on an Nvidia RTX 4090, with day-one support in llama.cpp across Apple, AMD, Qualcomm and Nvidia chips. d1-3B reads text and images; the smaller d1-omni adds audio. The company reports a score of 48.57 on its own Decision Index, ahead of every model under 10 billion parameters it tested and level with a decision model twelve times d1-3B's size. Those are Liquid's figures, measured on Liquid's benchmark, so a degree of caution is in order until outside researchers run the open checkpoints themselves.

The edge angle

The pitch is squarely at devices that cannot phone home to a data centre for every decision. A 600-million-parameter model that answers in milliseconds can sit on a phone, a camera or a factory sensor and never touch the cloud. That is a useful counterpoint to the running complaint that the industry's idea of open keeps getting heavier. Only this week we noted that many of the new open models will not run on your laptop, even as community projects strain to squeeze a 125-billion-parameter model onto a single gaming GPU. Liquid is going the other way, and the cheaper, smaller end of the market is getting more interesting for it, much as Anthropic's cut-price Haiku 5.5 did last week. The weights ship under Liquid's LFM Open License, which carries conditions on commercial use, so read the model card before you build on it.