The Pulse
Liquid AI Releases Open d1 Models With 16ms Jetson Decisions
Liquid AI released d1-3B and experimental d1-omni-600M as open-weight multimodal decision models on October 7, 2026. The company reports d1-3B answers a single question in 16 milliseconds on an NVIDIA Jetson AGX Thor, while noting that the
AI.info Team · Written with AI ·
16 milliseconds on a Jetson AGX Thor
Liquid AI says its new d1-3B model can return a decision in 16 milliseconds on an NVIDIA Jetson AGX Thor. The company released d1-3B and the experimental d1-omni-600M as open-weight models on October 7, positioning them for structured decisions on edge hardware as well as larger systems.
Unlike a generative model that writes an answer token by token, a d1 decision model processes input in one forward pass and returns an answer for a defined question. Developers can ask for a yes-or-no result, a choice among labels or a score on a scale. The models accept text and images; d1-omni-600M also accepts text paired with audio.
Two models, different hardware trade-offs
The 3B model builds on Liquid AI’s LFM2.5-VL-3B vision-language model. The 600M experimental model uses the LFM2.5-Encoder-350M backbone with added vision and audio encoders. Liquid AI describes d1-omni-600M as an early research release still in development, and provides no speed measurements for it.
For d1-3B, the company reports one-question response times of 16 milliseconds on Jetson AGX Thor, 26 milliseconds on Jetson AGX Orin 64 GB and 50 milliseconds on Jetson Orin Nano. A 384-pixel image takes 35 milliseconds on Thor and 202 milliseconds on Orin Nano, according to the same tests. The release says three questions take 20 milliseconds on Thor, compared with 16 milliseconds for one.
Benchmark results have a clear boundary
Liquid AI tested both models on seven public datasets covering reading comprehension, toxicity detection, intent classification, medical questions and cross-lingual understanding. It reports mean scores of 82.9 for d1-3B and 78.4 for d1-omni-600M. In the company’s comparison table, Decider 4B scores 81.1 and Decider 2B scores 77.1, respectively.
Those results do not establish how well the models make decisions from images or audio. Liquid AI says it is not publishing vision or audio benchmark scores: its cited Decision Index version includes only a private vision split, and it describes audio decision benchmarks as an open problem. The reported dataset scores therefore cover a different evaluation from the models’ claimed multimodal capability.
The company also reports a 48.57 score for d1-3B on Decision Index 0.2.1, which it identifies as the highest result among models under 10 billion parameters. It says that score beats Decider 35B-A3B’s 47.11. The release does not make that index score interchangeable with the seven-dataset mean; they are separate comparisons.
Weights are available for local use
Both models are available to download from Hugging Face, and Liquid AI points developers to a set of interactive demos in its System One Arcade. The release’s example requires Transformers version 5.14 or later, along with PyTorch, torchvision and Pillow. Loading the model also requires setting trust_remote_code=True, a deployment detail developers should account for when reviewing and running model code.
The release makes an edge-focused decision system available beyond Liquid AI’s hosted API, but its strongest performance claims are specific: latency measurements are for d1-3B, while d1-omni-600M remains experimental and lacks published speed tests. The company also has not supplied public vision or audio decision benchmarks for either model.