Liquid AI Releases Two Open-Weight Decision Models That Answer in Under 50 Milliseconds on Edge Hardware
The d1-3B and d1-omni-600M models skip word-by-word generation and return structured answers in a single calculation, fast enough to run on a $250 Jetson Orin Nano board.
Key points
- Liquid AI's d1-3B scores 48.57 on the Decision Index 0.2.1, beating every competing model under 10 billion parameters, including the 35-billion-parameter Decider 35B-A3B (47.11).
- D1-3B answers a single question in 16 milliseconds on a Nvidia Jetson AGX Thor, a small computer designed for on-device AI rather than cloud computing.
- The smaller companion model, d1-omni-600M, accepts text, images, and audio while using roughly a quarter of the parameters of the 2-billion-parameter Decider it outperforms.
- Both models are open-weight, meaning anyone can download and inspect the underlying numbers, and are available on Hugging Face today.
Most AI models live in a data centre and send your request across the internet before answering. Liquid AI's new d1 models are built to do the opposite: sit on a small computer attached to a camera or a sensor and decide things locally in the time it takes to blink.
D1-3B and the experimental d1-omni-600M are what the company calls "decision models." A standard AI generates responses word by word. A decision model skips that entirely: give it a question and a list of valid answers, and it returns a structured choice in a single mathematical pass through its weights, the stored numerical knowledge baked into the model during training.
What does 16 milliseconds actually mean?
For a factory vision system, 16 milliseconds means a computer that fits in your hand is keeping up with a fast conveyor belt far more easily than a human inspector could. Liquid benchmarked d1-3B alongside Nvidia across four devices. On a desktop GPU, the RTX 4090, one question takes 8 milliseconds. On the Jetson Orin Nano, it's 50 milliseconds. Three questions together take only 1.3 times as long as one, which matters for anyone running multiple checks in parallel.
| Device | One question | Three questions | Image (384px) |
|---|---|---|---|
| Nvidia RTX 4090 | 8 ms | 21 ms | 17 ms |
| Jetson AGX Thor | 16 ms | 20 ms | 35 ms |
| Jetson AGX Orin 64 GB | 26 ms | 35 ms | 83 ms |
| Jetson Orin Nano | 50 ms | 73 ms | 202 ms |
On seven public text benchmarks covering reading comprehension, toxicity detection, intent classification, and medical question-answering, d1-3B posts a mean score of 82.9, above the 4-billion-parameter Decider 4B at 81.1. D1-omni-600M scores 78.4, above the Decider 2B at 77.1 despite being roughly a quarter its size.
Who would actually use this?
Anyone who needs a fast yes-or-no or a priority score without running a full chatbot. The release blog shows a customer-service ticket being routed to billing or fraud teams in one pass. That pattern fits a drone classifying what it sees below, a medical device flagging an abnormal reading, or a retail shelf camera checking stock levels just as well.
D1-omni-600M adds audio on top of text and images, opening it to voice-driven edge devices, though Liquid flags it as an early research release still under development.
We've followed Liquid AI since our first story on 28 July 2026, and the pace has picked up: the d1 models are built directly on the LFM2.5-VL-3B backbone that featured in our 24 September coverage of Liquid's vision speed work. Knowing the lineage helps set realistic expectations: Liquid's benchmarks here draw on a private vision split it controls, so external replications will matter more than the headline numbers.
The open-weight release is what I'd watch closely. Closed models ask you to trust a company's server; open weights let an engineer audit, fine-tune, or deploy behind a firewall. A decision model this fast at this size is a meaningful step forward, but the benchmark story isn't finished until someone outside Liquid runs the same tests.
Both models are downloadable from Hugging Face now.
