Liquid AI Open-Sources Its Decision Models, Betting Edge Hardware Wants Answers Instead of Tokens

Liquid AI released two open-weight models on October 7 — d1-3B and an experimental d1-omni-600M — extending a decision-model product line it introduced as a hosted API just two days earlier. Both are available on Hugging Face under what the company describes as an open-weight license letting anyone download, fine-tune, and deploy the models "without restrictions." The release is notable less for its size than its category: decision models don't generate text. They take unstructured input — text, an image, audio, or some combination — plus one or more questions, and return a probability in a single forward pass, with no token-by-token generation at all.

Answers, not essays

Liquid AI's framing splits a decision model's output into three question types: yes/no questions that return a probability between 0 and 1, single-choice questions across many labels that return a probability per label, and scored questions on a scale, weighted by the probability of each level. A single request can bundle several questions about the same input, and because the model never generates free-form text, a text-only decision resolves in roughly 200 to 300 milliseconds using the original hosted d1 model, with the smaller open-weight d1-3B answering in under 50 milliseconds on every device the company measured and under 10 milliseconds on an Nvidia RTX 4090. The company tested the hosted d1 against GPT-6.1 Sol and Claude Opus 5.5 on six real applications — including support-ticket filtering, code search, document filing, web-agent navigation, and compacting tool output for coding agents, which it says cut token counts by 52 percent — and reported matching or beating GPT-6.1 Sol on four of the six, at a claimed 19x to 200x lower cost and faster latency on every task measured.

Why ship small, open versions a week later

The October 7 follow-up pushes the same architecture toward edge hardware rather than the cloud API Liquid AI launched with. d1-3B is distilled from the company's LFM2.5-VL-3B vision-language model and handles text-plus-image input; d1-omni-600M, built on LFM2.5-Encoder-350M, is lighter and more experimental, accepting either text-plus-image or text-plus-audio. On Liquid AI's own Decision Index benchmark, d1-3B scored 48.57 on the public split — ahead of every model under 10 billion parameters and roughly matching Decider 35B-A3B, a model the company says is about 12 times its size. Across seven general text benchmarks, d1-3B averaged 82.9 versus 81.1 for the similarly sized Decider 4B, while d1-omni-600M averaged 78.4 versus 77.1 for Decider 2B at roughly a quarter of the parameter count. The company says both models have day-one support in llama.cpp across Apple, AMD, Qualcomm and Nvidia hardware, and demonstrated them with ten live-camera demo apps — gesture-controlled games and real-time content moderation among them — plus a hardware-in-the-loop navigation demo running on an Nvidia Jetson board inside Isaac Sim.

A crowded, fast-moving niche

Liquid AI is not alone in betting that narrow, non-generative classification is a distinct product category worth building hardware-optimized models for. Cloudflare released its own open-weight decision models, Clef and Clef-flash, on October 1, and AutoTrust AI's JEV-27B-VL has separately topped Hugging Face's global trending list as open decision models gain traction — both developments that predate Liquid AI's d1 launch by about a week. Liquid AI's own benchmark page links its Decision Index to a Hugging Face Space whose URL contains "jev," suggesting some shared lineage or borrowed tooling in how the field measures these models, though Liquid AI's materials don't name AutoTrust, Cloudflare, or any other decision-model vendor directly, comparing d1 only against its own earlier Decider-branded models.

What ties the three releases together is a shared wager: that as AI agents increasingly need cheap, fast, repeated judgment calls — is this ticket urgent, does this frame contain a defect, should this web agent click here — running a full generative model for every such call is wasteful overkill, and a market exists for small, specialized models that skip text generation entirely. Liquid AI's published numbers on cost and latency make that case most aggressively of the three, but none of the vendors has yet published results from an independent third party rather than its own benchmark suite, and Liquid AI's own roadmap — more decision models across a range of sizes, with open weights for future releases also planned on Hugging Face — suggests the company sees this as the opening stage of a release cycle rather than a one-off product launch.