Liquid AI d1 Models: Open Decision AI That Answers in 8 ms

Liquid AI has released open weights for its d1 decision models, a new kind of small AI model built to make choices rather than write text. Published on October 7, 2026, the d1-3B and d1-omni-600M models return yes/no probabilities, labels or rubric scores in a single forward pass, without generating any tokens. The payoff is speed: Liquid AI reports d1-3B answering a single question in 8 milliseconds on an RTX 4090 and 50 milliseconds on a compact Jetson Orin Nano.

• Two models: d1-3B (text and images, built on LFM2.5-VL-3B) and d1-omni-600M (text plus image or audio, built on LFM2.5-Encoder-350M and labeled an experimental checkpoint).
• Benchmarks (Liquid AI): d1-3B scores 48.57 on Decision Index v0.2.1, which Liquid AI says matches a 35B-parameter model about 12 times its size.
• Latency (Liquid AI): 8 ms on RTX 4090, 9 ms on AMD MI325X, 16 ms on Jetson AGX Thor, 30 ms on Apple M5 Pro and 50 ms on Jetson Orin Nano.
• Availability: open weights on Hugging Face with day-one llama.cpp support across NVIDIA, Apple, AMD and Qualcomm hardware.

What Is a Decision Model?

Large language models are generalists: they answer by producing text one token at a time. But a huge share of real AI work is not open-ended writing. It is a decision. Is this comment toxic? Does this image show a fallen package? Is the robot's path clear? Generating a paragraph to answer a yes/no question wastes time and compute.

Liquid AI's d1 models skip generation entirely. They read the input and output the decision directly, in one pass. That makes them a natural fit for content moderation, routing, quality checks, robotics and any edge device where every millisecond and every watt counts. We saw the same idea from a different team earlier this month with Strands Decider 2B, and d1 pushes it into multimodal territory.

How Fast Are the d1 Models on Edge Hardware?

Liquid AI published latency and throughput figures across a wide hardware range:

• Single-question latency for d1-3B: RTX 4090 8 ms, AMD MI325X 9 ms, Jetson AGX Thor 16 ms, Jetson AGX Orin 64GB 26 ms, Apple M5 Pro 30 ms, Jetson Orin Nano 50 ms.
• Packed throughput (64 states): 1,106 decisions per second on the MI325X, 475 on the RTX 4090, 262 on Jetson AGX Thor and 38 on Jetson Orin Nano.

Those Jetson numbers are the exciting part for robotics builders. A 50 ms decision on an Orin Nano is fast enough for real-time camera checks on a small robot or smart camera, and Liquid AI built a hardware-in-the-loop demo with NVIDIA where d1-3B navigates an Isaac Sim environment from a Jetson.

Benchmark Results in Context

On Liquid AI's own Decision Index v0.2.1, d1-3B scored 48.57, which the company says is ahead of every model under 10 billion parameters and on par with Decider 35B-A3B at 47.11. On a mean of seven public text-classification tasks, d1-3B scored 82.9 and d1-omni-600M scored 78.4, compared with 81.1 for Decider 4B and 77.1 for Decider 2B. These are vendor-reported results, so independent testing will tell the fuller story, but the size-to-accuracy ratio is notable.

Where to Get d1 and How to Run It

Both models are on Hugging Face under the LiquidAI organization, with documentation at Liquid AI's docs site. Day-one llama.cpp support, including NVFP4, means you can run them on common local AI setups without waiting for third-party ports. Liquid AI describes them as open-weight models you can download, fine-tune and deploy, and RuntimeWire reports the model cards list the LFM Open License v1.0, so check the license terms for commercial use. Ten live-camera demos are available in a Hugging Face Space.

Liquid AI has been steadily building out its small-model lineup, from the LFM2.5-VL-3B on-device vision model that d1-3B is built on to new encoders. Decision models are a smart next step: small, fast and purpose-built for the moments when an edge device simply needs the right answer now. More in our AI coverage.

Sources: Liquid AI — d1: open decision models — October 7, 2026; Hugging Face Blog — Liquid AI open d1 — October 7, 2026; RuntimeWire — Liquid AI open d1 decision models for the edge — October 7, 2026.