Liquid AI Unveils open d1: Ultra-Fast Edge Decision Models with Single-Pass Multimodal Inference
Liquid AI 推出 open d1 邊緣決策模型：無需生成 Token、單次前向傳播即刻做出多模態決策
Liquid AI has introduced open d1, a family of edge-focused decision models built on Liquid Foundation Models (LFMs). Instead of generating tokens step-by-step, these models run in a single forward pass to output structured decisions directly, drastically reducing latency. d1-3B supports text and images, answering in under 50ms on edge devices like NVIDIA Jetson. d1-omni-600M is an early research release supporting text-image or text-audio modalities, beating Decider 2B with just a fraction of its parameters.
Key points
Tokenless Single-Pass Inference
Unlike generative models, d1 outputs structured decisions in a single forward pass, completely bypassing token generation latency.
High-Performance Edge Deployment
d1-3B answers single questions in under 50ms on NVIDIA Jetson and Apple M5 Pro, and under 10ms on discrete GPUs.
Multimodal Decision Processing
d1-3B supports text and image inputs, while the 600M-parameter d1-omni handles text/image or text/audio modalities.
How it works
| d1-3B | d1-omni-600M | |
|---|---|---|
| Parameter Size | 3B (30億) | 600M (6億) |
| Base Backbone | LFM2.5-VL-3B (僅解碼器) | LFM2.5-Encoder-350M (雙向編碼器) |
| Supported Modalities | 文字 + 影像 | 文字 + 影像 或是 文字 + 語音 |
| Mean Benchmark Score | 82.9 | 78.4 |
| Status | 正式發布 / 已測試邊緣速度 | 早期研究版本 (Early Research) |
Why it matters
This development showcases the massive potential of non-generative 'decision models' on edge hardware. By bypassing expensive token generation steps and solving classification, scoring, and routing tasks in a single forward pass, d1 makes real-time, low-latency, and privacy-centric local multimodal AI applications viable on low-power devices without relying on costly cloud GPUs.
Who it affects
- AI Developer
- AI Researcher
- Enterprise Leader
- Product Manager
How to use it
- 1Smart Routing & Intent Classification: Automatically triage user requests or voice commands to specific teams like billing or support.
- 2On-Device Content Moderation: Instantly analyze local text or images for policy violations and toxicity on edge devices.
- 3IoT & Robotic Control: Execute millisecond-level local control decisions based on real-time vision or audio sensor inputs.
Limitations & caveats
- d1-omni-600M is currently an early research release with no benchmarked inference speed numbers reported yet.
- Being decision-only models, they cannot perform generative tasks such as drafting textual responses or free-form chat.
- Lack of public audio decision benchmarks limits standard evaluation of the model's multimodal performance.
Related
NVIDIA cuPhoton: Accelerating End-to-End Scientific Image Analysis on GPUs
NVIDIA cuPhoton 開源工具套件：加速天文與科學影像端到端 GPU 分析
NVIDIA cuPhoton is an open-source CUDA-X toolkit providing GPU-accelerated building blocks that eliminate CPU bottlenecks in end-to-end scientific image processing.
NVIDIA DIN Deploy: Building High-Performance Local AI Apps with C++ and TensorRT RTX
NVIDIA 推出 DIN Deploy：用 C++ 與 TensorRT RTX 打造高效能地端 AI 應用程式
NVIDIA's open-source DIN Deploy combines ONNX Runtime and TensorRT RTX, providing C++ samples to help developers run ASR, segmentation, and image generation models locally with hardware acceleration on Windows and Linux.
Designing AI-Native Software: Lessons from NVIDIA TensorRT Model Connect
打造 AI 原生專案：NVIDIA TensorRT Model Connect 的開發啟示
NVIDIA shares architectural insights from building TensorRT Model Connect, demonstrating how to design software around coding agents using model-family isolation, reversible changes, and automated validation.