HOT RADAR
llmHEAT 0.0°

AI CURATED HOT STORY · 10/8/2026

Liquid AI Open-Sources d1-3B and d1-omni-600M: Achieving 8ms Edge Multimodal Decision-Making via LFM2.5-VL Architecture

2 reports archived2 independent sourcesupdated 10/9/2026, 14:29:07
Synthesis & Latest Updates
2 Sources Cross-Validated

Liquid AI has officially open-sourced its d1 series decision models, featuring the d1-3B and the experimental multimodal d1-omni-600M. This release targets the latency bottleneck in edge computing for real-time agent applications, with d1-3B redefining performance benchmarks for lightweight decision models through ultra-low inference times and robust multimodal capabilities.

Cross-verified reports indicate that d1-3B is trained on the LFM2.5-VL-3B architecture with 3.12 billion parameters. On an NVIDIA RTX 4090, it achieves a single text judgment in just 8 milliseconds and processes 384-pixel images in approximately 102 milliseconds; even on constrained edge devices like the Jetson Orin Nano, response times remain under 50 milliseconds. In the Decision Index v0.2.1 benchmark, the model scored 48.57, surpassing competitors ranging from 4B to 35B parameters, establishing itself as the strongest decision model under 10B parameters. Additionally, d1-omni-600M supports combined text, visual, and audio inputs, offering broader perception capabilities for resource-limited environments.

This initiative significantly lowers the barrier to deploying high-performance multimodal agents, enabling robots, autonomous driving assistants, and IoT devices to execute complex decisions locally with minimal latency. With model weights fully open-sourced on Hugging Face, the developer community will accelerate the exploration of new paradigms for real-time edge interaction, driving the ecosystem shift from cloud-based large models to efficient on-device intelligence.

LATEST/Liquid AI has fully released the model weights and technical documentation for d1-3B and d1-omni-600M on Hugging Face. The community is currently replicating their Decision Index v0.2.1 benchmark results and exploring the integration of d1-omni-600M into more low-power hardware prototypes to verify stability and scalability in real-world multimodal interaction scenarios.

Coverage timeline

Total 2 reports · Latest first
  1. IT之家 · 智能时代T2·78 pts
    • The d1-3B model, trained on LFM2.5-VL-3B with 3.12 billion parameters, scores 48.57 on Decision Index v0.2.1, claiming superiority over all sub-10B models.
    • Exceptional inference speed: d1-3B takes only 8ms for single text queries and 102ms for 384-pixel images on an RTX 4090; it maintains sub-50ms response times even on edge devices like Jetson Orin Nano.
    • d1-omni-600M is Liquid AI's first experimental multimodal decision checkpoint, supporting combined text, image, and voice inputs for resource-constrained scenarios.