Open Source AI

Liquid AI releases open-weight decision models for fast edge-device responses

Hugging FaceOct 7, 2026

Liquid AI released d1-3B and experimental d1-omni-600M, reporting strong decision benchmarks and millisecond-scale inference on edge hardware.

Liquid AI released two open-weight models in its d1 decision-model family on October 7, 2026: d1-3B and the experimental d1-omni-600M. The company describes them as models for producing structured decisions from inputs in a single forward pass, rather than generating a sequence of text tokens.

The models differ in size and input capabilities. d1-3B, built on Liquid AI’s LFM2.5-VL-3B, accepts text and images. The smaller d1-omni-600M is based on LFM2.5-Encoder-350M and can process text with either images or audio; Liquid AI says it remains an early research release. Both models are available as open weights.

Liquid AI reports a score of 48.57 for d1-3B on Decision Index 0.2.1, placing it ahead of the 4B and 9B models in that comparison and Decider 35B-A3B, which scored 47.11. On a separate set of seven public datasets covering tasks such as reading comprehension, intent classification, and medical question answering, the company reports mean scores of 82.9 for d1-3B and 78.4 for d1-omni-600M. It says the latter exceeded Decider 2B’s 77.1 score while using one-quarter as many parameters. These results are company-reported; the release does not provide vision or audio benchmark scores.

The release also reports inference measurements for d1-3B across edge devices. A single question took 16 milliseconds on NVIDIA Jetson AGX Thor, 26 milliseconds on Jetson AGX Orin 64 GB, and 50 milliseconds on Jetson Orin Nano. The model also ran on Apple M5 Pro, NVIDIA RTX 4090, and AMD MI325X systems. Liquid AI says three questions took 1.3 times as long as one on the measured edge devices, with the AGX Thor processing them in 20 milliseconds.

The reported results position the models for applications that need quick, structured responses from text, images, or audio, including deployments on local hardware. Liquid AI provides example usage for d1-3B and says both releases can be tried through Hugging Face model listings and demos.

Industry Impact

Structured AI decisions move closer to the edge

Edge computing, AI software, and on-device applications

Overview

Liquid AI’s release gives software developers and ML teams open-weight models to evaluate for applications that need structured decisions from text, images, or audio on local hardware. The company reports millisecond-scale d1-3B inference on several edge devices and competitive scores on text-focused decision benchmarks, but these results are company-reported and no vision or audio benchmark scores are provided. In the near term, teams can test latency and accuracy against their own workloads; longer term, local inference could support responsive features without relying on a remote model for every decision.

Key Impact Points

d1-3B returned a single question’s result in 16–50 milliseconds across three reported Jetson edge devices.

Open weights let developers assess deployment fit and adapt integrations, while the smaller d1-omni-600M remains an experimental release.

Benchmark claims are promising but do not establish performance for vision or audio tasks.

Opportunities

Local decision features for edge products

Edge devices, embedded applications, and AI software

Product teams can evaluate these open-weight models for features that convert text, images, or audio into structured decisions on local hardware. Potential applications include responsive classification, intent handling, and question-answering workflows, although the article does not validate specific deployments or provide vision and audio benchmark scores. Developers should test end-to-end quality, latency, memory use, and reliability on their target devices before committing. The reported measurements and benchmark results offer a starting point, not a guarantee of production performance. The release provides no pricing or commercial terms, so any revenue case would depend on the value of the resulting product or workflow rather than a stated model business model.

On-device structured decision workflows
Problem: Products that need timely structured responses may be constrained by remote inference latency or connectivity requirements.
Solution: Evaluate Liquid AI’s open-weight decision models for local inference on supported edge hardware, matching inputs and tasks to each model’s capabilities.
Target Users: Developers and product teams building edge-device applications that process text, images, or audio.
Business Value: Potentially enable responsive local features; validate quality, latency, and device fit with workload-specific testing.
Revenue Potential: No pricing or revenue model is stated. Commercial value could come from differentiated product features or workflow improvements if deployments prove useful.

Reference link

https://huggingface.co/blog/liquidai/open-d1

Image Source Link

https://blogs.nvidia.com/wp-content/uploads/2026/05/26gtc-tpe-keynote-GM_05466-1-1280x720.jpg

Community Discussion (0)

Loading discussions...