Liquid AI open-sources d1 multimodal decision models: 16ms response on Jetson AGX Thor

On October 7, 2026, Liquid AI published an article titled "Multimodal open d1 decision models for the edge" on the Hugging Face blog, announcing two open-source models in its d1 decision model family: d1-3B and the experimental d1-omni-600M. The article was co-authored by Liquid AI team members Aurelien Lac, Fernando Fernandes Neto, Edoardo Mosca, Maxime Labonne, Leonie Monigatti, and others. These two models target fast structured decision-making in edge scenarios, in stark contrast to traditional generative models that produce tokens continuously: decision models give answers directly in a single forward pass.

Architecturally, both models are built on Liquid AI's self-developed Liquid Foundation Models (LFMs). d1-3B is trained on the latest vision-language model LFM2.5-VL-3B, a decoder-only VLM that accepts text and image inputs; d1-omni-600M is trained on the bidirectional encoder LFM2.5-Encoder-350M, with additional vision and audio encoders for three modalities, accepting "text+image" or "text+audio" inputs, and is in an ongoing experimental phase.

The article states: "d1-3B answers a question in 16 ms on an NVIDIA Jetson AGX Thor, 26 ms on a Jetson AGX Orin, and 50ms on a Jetson Orin Nano", highlighting its low-latency edge deployment characteristics. In business metrics, d1-3B scored 48.57 on Decision Index 0.2.1, becoming the best sub-10B-parameter decision model, surpassing all 4B and 9B models as well as the larger Decider 35B-A3B (47.11).

In benchmarks, Liquid AI evaluated both models on seven public datasets (covering reading comprehension, toxicity detection, intent classification, medical Q&A, and cross-lingual understanding tasks), with d1-3B averaging 82.9, the highest in the table, beating Decider 4B; d1-omni-600M scored 78.4, surpassing Decider 2B (77.1) with just one quarter of the parameters. Specifically, d1-3B achieved the best results on SQuAD 2.0 (83.3) and PubMedQA (68.3), while d1-omni-600M led on Civil Comments (95.8) and PAWS-X (79.5). The article also verified that d1-3B retains the visual capability of its LFM2.5-VL-3B backbone, and d1-omni-600M also handles all three modalities. However, since Decision Index v0.3 contains only private visual segmentation data and audio decision benchmarks remain an open problem, they did not publish any visual or audio benchmark scores.

On speed, Liquid AI partnered with NVIDIA to evaluate d1-3B on NVIDIA GeForce RTX 4090, NVIDIA Jetson AGX Thor, Jetson AGX Orin 64 GB, and Jetson Orin Nano. For edge inference, d1-3B answers a single question in under 50ms on all tested devices; asking three questions takes only 1.3x the time of one, with AGX Thor going from 16ms to 20ms. For GPU inference, d1-3B answers questions in under 10ms and processes 384px images in under 18ms on both platforms: 8ms per question, 17ms for 384px image, 475 packed states per second on RTX 4090; 9ms per question, 18ms for 384px image, 1,106 states per second on AMD MI325X.

On commercial impact, both models are released as open-weight and available on Hugging Face at huggingface.co/LiquidAI/d1-3b and huggingface.co/LiquidAI/d1-omni-600M; users can also try ten demos in the System One Arcade space hosted by Liquid AI on Hugging Face Spaces. Overall, this update reinforces Liquid AI's position in edge AI, making fast, low-latency multimodal structured decision-making possible on Jetson and other edge hardware, providing substantial momentum for the open-source community and end-application developers.