On October 7, 2026 Liquid AI published two open decision models in its d1 family: d1-3B and the experimental d1-omni-600M. These decision models are designed to provide a single-pass, structured answer rather than token-by-token generation.
Model design and modality support
d1-3B is built from the LFM2.5-VL-3B backbone, a decoder-only Vision-Language Model, and accepts text and images. d1-omni-600M is trained from LFM2.5-Encoder-350M, a bidirectional encoder, and adds dedicated vision and audio encoders; it accepts text+image or text+audio inputs. Liquid AI notes that d1-omni-600M is an early research release under further development.
Benchmarks and rankings
According to Liquid AI, d1-3B achieved the top Decision Index v0.2.1 result under 10B parameters with a score of 48.57, ahead of all 4B and 9B models and Decider 35B-A3B (47.11). On seven public decision datasets spanning reading comprehension, toxicity, intent classification, medical QA and cross-lingual tasks, the reported mean scores were: d1-3B 82.9, d1-omni-600M 78.4, Decider 2B 77.1 and Decider 4B 81.1.
Selected dataset results reported by Liquid AI include SQuAD 2.0 (d1-3B 83.3, d1-omni-600M 74.0), PubMedQA (d1-3B 68.3, d1-omni-600M 61.3) and XNLI (d1-3B 85.6, d1-omni-600M 74.7). The team also validated that d1-3B retained vision capabilities from its backbone and that d1-omni-600M handles the three modalities; no public vision or audio decision benchmarks were reported because Decision Index v0.3 includes a private vision split and audio decision benchmarks remain an open problem.
Performance on edge and GPU
Liquid AI reported edge inference timings for d1-3B across NVIDIA hardware: one question in 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin (64 GB) and 50 ms on Jetson Orin Nano. Throughput figures cited include 262 requests/s on AGX Thor and 38 requests/s on Orin Nano. On GPU, d1-3B answered a single question in under 10 ms: 8 ms on an NVIDIA RTX 4090 and 9 ms on an AMD MI325X, with reported throughputs of 475/s and 1,106/s respectively.
Both models are released as open-weight and are available on Hugging Face. Reported software requirements include transformers>=5.14; installation guidance given by Liquid AI is: pip install "transformers>=5.14" torch torchvision pillow. The models provide their own loader code and the publisher recommends loading with trust_remote_code=True.
Original source: Hugging Face Blog