Liquid AI has released Open d1, consisting of two open-weight multimodal decision models named d1-3B and d1-omni-600M, designed for real-time decisions on the NVIDIA stack with zero output tokens. Unlike traditional generative models that write answers token by token, these decision models take a state and a set of named questions, returning calibrated and typed answers in a single forward pass. The models are available on Hugging Face with day-one llama.cpp support under the LFM Open License v1.0, which permits free commercial use below $10 million in annual revenue.
The d1-3B model features 3.12B parameters, reads text and images, and is built from a decoder-only vision-language model using a SigLIP2 NaFlex vision encoder and a 32,768-token context. Meanwhile, the d1-omni-600M model contains 587M parameters, reads text with an image or audio, and utilizes a bidirectional encoder paired with a 17-layer FastConformer audio encoder for clips up to 30 seconds. On the Decision Index v0.2.1 benchmark, d1-3B scores 48.57, outperforming every model under 10B parameters, while d1-omni-600M averages 78.4 across seven public text benchmarks.
These models are optimized for tasks such as routing, moderation, intent classification, reranking, LLM-as-a-judge scoring, and agent guardrails rather than standard chat applications. Testing shows d1-3B processes a single question in 8 ms on an RTX 4090 and 16 ms on a Jetson AGX Thor. This release matters because it provides developers with highly efficient, low-latency multimodal decision models capable of instant inference without parsing text generation.