Liquid AI Opens d1-3B and d1-omni-600M Decision Models
TL;DR
- d1-3B matches Decider 35B-A3B on Decision Index v0.2.1 at 12x smaller, challenging the assumption that classification scales with generative model size.
- Zero output tokens remove the parse-LLM-JSON step; the model returns typed calibrated answers directly from probability space in one forward pass.
- At 16ms on Jetson AGX Thor, d1-3B puts sub-20ms multimodal classification within reach for robotics and edge deployments without cloud dependency.
Liquid AI's new d1-3B scores 48.57 on the Decision Index v0.2.1, a benchmark the company controls, which it says puts the model "ahead of every model under 10B and on par with Decider 35B-A3B, a decision model 12x its size." Its smaller sibling d1-omni-600M scores 15.95 on the same split.
The models do not generate text. "Unlike our generative Liquid Foundation Models (LFMs), our d1 decision models don't produce tokens. Instead, they produce an answer in a single forward pass," Liquid AI writes in the release post. d1-3B inherits from LFM2.5-VL-3B and takes text plus images. d1-omni-600M is built on LFM2.5-Encoder-350M with added vision and audio encoders.
The pitch is latency. A single question runs in 8 ms on an NVIDIA RTX 4090, 9 ms on an AMD MI325X, 16 ms on a Jetson AGX Thor, 26 ms on a Jetson AGX Orin, and 50 ms on a Jetson Orin Nano. A 384-pixel image lifts those to 17 ms and 202 ms at the two extremes. The use cases Liquid AI floats are "gesture-controlled games" and "live content moderation" running frame-by-frame on camera input.
Both models are on Hugging Face with native llama.cpp support across Apple, AMD, Qualcomm, and NVIDIA platforms, described as open-weight and free to "download, fine-tune, and deploy without restrictions." The post does not name the specific license text that governs that promise. It lands in a busy week for small-model releases: Perplexity shipped its pplx-embed-v2-late embeddings the same day, and llama.cpp itself picked up GLM-5.3-Flash last week.
What others are reporting
-
AICoder Read →
Anchors the release in benchmark numbers: d1-3B first among sub-10B on Decision Index v0.2.1 and on par with Decider 35B-A3B at 12x smaller.
d1-3B scores 48.57 on Decision Index v0.2.1, first among sub-10B models and on par with the 12x larger Decider 35B-A3B
-
DEV Community Read →
Hands-on CPU testing of the 600M model with documented failure modes: 0.907 confidence on nonsense input, practical guardrail code, and median 208ms latency.
A decision model does not write text. You hand it a state and a set of named questions, and it returns typed, calibrated answers in one forward pass.
-
Eigen Radar Read →
Flags audio evaluation as an open research problem, framing this as a usable but incomplete release rather than a finished product.
They take a supplied question and return structured decisions in one forward pass.
-
RuntimeWire Read →
Edge-first framing: covers the single-forward-pass architecture and structured output design as a direct replacement for generative inference pipelines.
Liquid AI's decision models instead process a state in a single forward pass and return structured answers.
-
Tech AI Magazine Read →
Highlights both models returning calibrated typed answers in a single forward pass, with clean coverage of the multimodal scope across text, image, and audio.
Both models return calibrated, typed answers in a single forward pass.
-
Superpower Daily Read →
Consumer-accessible framing of the no-token-generation design; covers multi-question single-pass capability for structured decision workloads.
The models can answer several structured questions about one input in a single pass without generating answer tokens.
Originally reported by liquid.ai
Read the original article →Original headline: Liquid AI Open-Weights d1-3B and d1-omni-600M, Decision Models With Zero Output Tokens