Liquid AI has released d1-3B and d1-omni-600M, two open-weight decision models optimized for real-time inference on edge hardware and data centers. These models support text, image, and audio inputs and are designed to perform tasks like content moderation and environment navigation without generating tokens.
The shift from generative token production to single-pass decision logic significantly reduces the computational overhead for real-time video workflows. By achieving sub-50ms latency on low-power edge hardware like the Jetson Orin Nano, these models allow streaming platforms to move content moderation and gesture-based navigation from the cloud to the device. This reduces bandwidth costs and improves privacy for interactive streaming applications. As the industry moves toward multimodal interfaces, the ability to process audio and visual inputs simultaneously on-device will be a critical differentiator. Watch for the development of standardized audio decision benchmarks to validate performance across competing multimodal architectures.
Liquid AI's d1-3B is now available on Hugging Face, where it is positioned as a single-pass decision model that returns calibrated, typed answers with zero output tokens. The model card specifies that d1-3B is built on LFM2.5-VL-3B and accepts text, JSON, images, or mixed states as input, making it suitable for routing, triage, moderation, intent classification, and LLM-as-a-judge scoring pipelines. Liquid AI published the d1-3B model card on Hugging Face with full latency benchmarks across NVIDIA Jetson and desktop hardware, including a packed throughput figure of 262 decisions per second on the Jetson AGX Thor.
The Decision Index v0.2.1 benchmark that d1-3B leads is a public evaluation framework for decision models that Liquid AI introduced alongside the d1 family. On this index, d1-3B scores 48.57 on the public split, surpassing every model under 10 billion parameters and matching Decider 35B-A3B, which is 12 times larger. The model also posts the highest toxicity detection score in its comparison table at 95.8 on Civil Comments and 79.5 on PAWS-X paraphrase identification, both of which are directly relevant to content moderation use cases in streaming workflows. Liquid AI has made both d1-3B and d1-omni-600M open-weight, meaning developers can download, fine-tune, and deploy without restrictions.
NVIDIA's Jetson platform serves as the primary edge deployment target for Liquid AI's decision models, with the company reporting day-one support for llama.cpp across Apple, AMD, Qualcomm, and NVIDIA hardware with NVFP4 quantization. The collaboration between Liquid AI and NVIDIA extends beyond benchmarking: the two companies demonstrated d1-3B navigating an environment in Isaac Sim with the model served on a Jetson in a hardware-in-the-loop setup, showing the model's applicability to embodied AI tasks alongside its content moderation and classification capabilities. This dual positioning, spanning both streaming content workflows and robotics, reflects Liquid AI's strategy of building a single decision-model architecture that serves multiple real-time inference markets.
Liquid AI has launched the d1-3B and d1-omni-600M open-weight decision models, which bypass token generation to provide single-pass inference. By achieving sub-50ms latency on edge hardware, these models enable streaming platforms to perform content moderation and gesture-based navigation locally, significantly reducing bandwidth costs and enhancing privacy for interactive video applications.
The d1-3B model uses single-pass decision logic to bypass token generation, achieving sub-10ms latency on high-end hardware. It is designed for tasks like content moderation, intent classification, and routing, matching the performance of models 12 times its size while reducing computational overhead.
Yes. Liquid AI reports inference speeds of 16 ms on NVIDIA Jetson AGX Thor and 50 ms on Jetson Orin Nano, making them suitable for edge deployments in streaming and robotics.
The d1-omni-600M is an experimental checkpoint within the d1 family that supports multimodal inputs, including text, image, and audio, allowing for more versatile real-time processing.
Liquid AI has released both d1-3B and d1-omni-600M as open-weight models, which are available for download and fine-tuning on Hugging Face.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source
