Liquid AI has released Open d1, featuring two open-weight multimodal decision models: d1-3B (text and vision) and d1-omni-600M (text with image or audio). Unlike generative LLMs, these decision models evaluate states against specific questions in a single forward pass and return probabilities without generating output tokens, targeted at real-time inference on NVIDIA hardware.

Both checkpoints are available on Hugging Face under the LFM Open License v1.0, which allows free commercial use for companies making under $10 million in annual revenue. The models feature day-one support for Transformers and llama.cpp, targeting applications like routing, moderation, intent classification, and guardrails.

On the Decision Index benchmark, d1-3B achieved a score of 48.57, outperforming several larger sub-10B models. Inference testing demonstrated low end-to-end latency, with d1-3B recording single-question response times as fast as 8 ms on an NVIDIA RTX 4090 GPU.

Why it matters

  • Offers a low-latency, zero-token alternative to generative LLMs for classification, routing, and guardrail tasks.

  • Enables ultra-fast edge and real-time AI deployments on standard NVIDIA hardware like Jetson and RTX.

Source: marktechpost.com