🔍 Read the full analysis: 5 Ways Multimodal Open D1 Decision Models Support Edge Computing on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Liquid AI released d1-3B and the experimental d1-omni-600M, open-weight models designed to return structured decisions in a single forward pass. The company reports benchmark and latency results for d1-3B, but independent evaluations and published vision and audio scores are not included.
Liquid AI has released two open-weight decision models, d1-3B and d1-omni-600M, designed to classify, score and answer tasks with a structured result in a single forward pass. The release targets uses including edge computing, where processing on or near a device can reduce response time or limit the need to send data elsewhere; performance figures in the announcement are company-reported and have not been independently replicated in the material provided.
The company says the models are built on its Liquid Foundation Models and are intended to return a decision rather than generate a sequence of conversational text. Potential tasks include routing customer requests, estimating urgency and answering questions about images. d1-3B is based on LFM2.5-VL-3B and accepts text and images. The smaller d1-omni-600M is built on LFM2.5-Encoder-350M with vision and audio encoders, and supports text paired with an image or audio.
On seven public datasets covering reading comprehension, toxicity detection, intent classification, medical question answering and cross-lingual understanding, Liquid AI reports mean scores of 82.9 for d1-3B and 78.4 for d1-omni-600M. The company’s comparison lists Decider 4B at 81.1 and Decider 2B at 77.1. The results vary by dataset: d1-3B scored below Decider 4B on BoolQ, MASSIVE intent and XNLI. The reported averages describe this selected evaluation set, not every task a developer might deploy.
For d1-3B, Liquid AI reports a response time of 16 milliseconds on an NVIDIA Jetson AGX Thor, 26 milliseconds on a Jetson AGX Orin 64 GB and 50 milliseconds on a Jetson Orin Nano for one question. The company also reports 8 milliseconds per question on an NVIDIA RTX 4090 and 9 milliseconds on an AMD MI325X. These measurements were conducted with NVIDIA, according to the release. The announcement provides no speed results for d1-omni-600M, which Liquid AI describes as an early research model still under development.
Decision Models at the Edge
The release offers developers a different model format for applications that need a bounded, structured response, such as a category, score or routing choice, rather than open-ended text. If the reported latency carries over to a specific product and hardware setup, a decision step could run close to where data is collected. That may be useful when network access is limited or when sending every request to a remote service is undesirable.
The smaller model could also be relevant where memory and compute are constrained. Liquid AI says d1-omni-600M has roughly a quarter of the parameters of Decider 2B while scoring higher on the company’s selected dataset mean. That is a comparison within the reported evaluation, not proof of superior performance across workloads. Teams would still need to test accuracy, reliability and resource use on their own tasks and devices.
As an affiliate, we earn on qualifying purchases.
How the Models Were Tested
The models are presented as part of Liquid AI’s Liquid Foundation Models family. The company’s evaluation covers seven public datasets: SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI and PAWS-X. The release also says it checked whether d1-3B retained vision capabilities from its vision-language backbone and whether d1-omni-600M handled supported modalities, but it does not publish vision or audio benchmark scores.
Liquid AI says Decision Index version 0.3 has a private vision split and that audio decision benchmarks remain an open problem. Its release describes d1-omni-600M as experimental. Both models are available as open weights on Hugging Face, with demos in the company’s System One Arcade Hugging Face Space. The stated release instructions call for Transformers version 5.14 or later and loading the models with their supplied code enabled.
“Best decision model under 10B on the Decision Index 0.2.1”
— Liquid AI
NVIDIA Jetson edge computing device
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Limits of the Published Results
The reported scores and latency figures come from Liquid AI’s release. The source material provides no independent evaluation, confidence intervals or enough testing detail to establish how closely the benchmarks match a particular deployment. Seven datasets measure selected capabilities; they do not establish accuracy, reliability or safety across all decision tasks. The release also does not publish vision or audio benchmark scores, or speed measurements for d1-omni-600M.
It remains unclear how either model handles ambiguous inputs, how often decisions would require human review, and how results change across different software configurations, input sizes and production workloads. The company’s measured latency should not be assumed to predict performance on other hardware or in a particular application. The experimental status of d1-omni-600M adds uncertainty about its future capabilities and operational characteristics.
AI structured decision models for edge
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing on Target Devices
Developers can download the open weights and compare the models with their own decision tasks, hardware and latency requirements. The next useful evidence would include independent benchmark results, reproducible test conditions, and published evaluations of the models’ vision and audio capabilities. For d1-omni-600M, additional results will also be needed to show how the experimental release performs as development continues.
Until then, the announcement supports treating the reported figures as an initial indication of potential, not a guarantee for production use. Teams considering deployment will need to measure end-to-end performance and decide how to check or escalate uncertain model decisions.
vision and audio AI models for edge
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Liquid AI release?
Liquid AI released d1-3B and d1-omni-600M, open-weight models designed to return structured decisions in a single forward pass. The company describes d1-omni-600M as experimental.
What does a single forward pass mean for these models?
Liquid AI says the models produce structured answers such as a classification or score, rather than generating a sequence of conversational tokens. The intended uses include routing requests and judging urgency.
How fast is d1-3B on edge devices?
Liquid AI reports one-question times of 16 milliseconds on Jetson AGX Thor, 26 milliseconds on Jetson AGX Orin 64 GB and 50 milliseconds on Jetson Orin Nano. These are company-reported measurements and may not match results in other setups.
Do the published tests establish performance for images and audio?
No. The release says the models support certain image and audio inputs, but it does not provide vision or audio benchmark scores. It also reports no speed measurements for d1-omni-600M.
Where can developers access the models?
Liquid AI says both models are available as open weights on Hugging Face and points users to demos in its System One Arcade Hugging Face Space. Its release instructions specify Transformers version 5.14 or later and use of the supplied code when loading the models.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
