🔍 Read the full analysis: Multimodal Open D1 Models: A New Approach To Edge Decision-Making on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Liquid AI released two open-weight models designed to return structured decisions in a single forward pass: d1-3B and the experimental d1-omni-600M. The company reports benchmark scores and sub-50-millisecond responses for d1-3B on tested edge devices, but independent evaluations and published vision and audio benchmark results are not included.
Liquid AI has released d1-3B and d1-omni-600M, open-weight models designed to return structured answers for decision tasks in a single forward pass. The company says the models are intended for uses such as classifying requests or judging urgency, and reports that d1-3B answered one question in 16 milliseconds on an NVIDIA Jetson AGX Thor; independent replication of the results is not provided.
The models are built on Liquid AI’s Liquid Foundation Models and are intended to classify, score or otherwise select an answer rather than generate a long sequence of text. Liquid AI describes possible tasks including routing a customer request to a team and answering a question about an image. Both models are available as open weights on Hugging Face.
Liquid AI says d1-3B is based on its LFM2.5-VL-3B vision-language model and accepts text and images. The smaller d1-omni-600M uses the LFM2.5-Encoder-350M bidirectional encoder with added vision and audio encoders; the company says it can process text paired with an image or audio. Liquid AI describes this model as an early research release still under development.
On seven public datasets, Liquid AI reports mean scores of 82.9 for d1-3B and 78.4 for d1-omni-600M. Its table lists scores of 81.1 for Decider 4B and 77.1 for Decider 2B. Results differ by dataset: d1-3B scores below Decider 4B on BoolQ, MASSIVE intent and XNLI. These are company-reported results from a selected evaluation, not independent confirmation of performance across decision tasks.
Why Edge Decision Speed Matters
For products that need to make a bounded decision close to where data is collected, latency and device requirements can affect whether a model is practical. Liquid AI’s reported tests place d1-3B on several devices, including the Jetson AGX Orin 64 GB at 26 milliseconds per question and the Jetson Orin Nano at 50 milliseconds. Those measurements give developers initial figures to compare with their own hardware and workloads.
The approach also targets tasks where a system needs a predefined output—such as a category, score or routing choice—rather than open-ended text generation. That could be useful for constrained applications, but the announcement does not establish that these models are more accurate, safer or less costly than alternatives in a specific deployment. Performance can vary with the input, software setup, device and task.
Liquid AI reports that three questions took 1.3 times as long as one on tested devices; on the AGX Thor, the reported time increased from 16 to 20 milliseconds. This may inform testing of grouped requests, but it is a measurement reported by the company, not a guarantee for production workloads. The smaller model’s reported dataset mean may also interest developers with tight hardware limits, though it does not settle how it performs on their own tasks.
NVIDIA Jetson AGX Orin development kit
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Models, Modalities and Test Scope
The release compares the models on seven public datasets spanning reading comprehension, toxicity detection, intent classification, medical question answering and cross-lingual understanding. The listed datasets include SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI and PAWS-X. A mean across these datasets summarizes only this selection; individual results vary and do not cover every type of decision.
Liquid AI says it checked that d1-3B retained vision capabilities from its vision-language backbone and that d1-omni-600M handled its supported modalities. However, the release provides no vision or audio benchmark scores. It says Decision Index version 0.3 includes a private vision split and that audio decision benchmarks remain an open problem. The reported Decision Index score for d1-3B—48.57 on version 0.2.1—is a separate company-cited result.
For speed testing, Liquid AI says it worked with NVIDIA and measured d1-3B on NVIDIA GPUs and Jetson devices, as well as Apple M5 Pro and AMD MI325X hardware. It reports 8 milliseconds per question on an NVIDIA RTX 4090 and 9 milliseconds on an AMD MI325X. The release does not report speed measurements for d1-omni-600M.
““Best decision model under 10B on the Decision Index 0.2.1.””
— Liquid AI
What the Published Tests Leave Open
The release does not include independent evaluations, confidence intervals or independent replication of its reported benchmark and speed results. It also does not provide enough information to determine how closely the test conditions match a particular deployment. The seven text-focused datasets do not establish accuracy, reliability or safety across all decision tasks.
Vision and audio performance is especially difficult to judge from the published material: Liquid AI gives no scores for either modality, and there are no speed results for d1-omni-600M. The announcement also does not explain how either model handles ambiguous inputs, how often decisions may need human review, or how performance changes under varied production workloads. Those questions remain for further evaluation.
Testing on Developer Hardware
Developers can download the open weights from Hugging Face and try demos in Liquid AI’s System One Arcade Hugging Face Space. The company’s instructions specify Transformers version 5.14 or later and say the models must be loaded with their supplied code enabled. Testing on relevant tasks and hardware will help determine whether the published figures apply to a particular use.
Further independent results, including evaluations of vision and audio tasks and performance for d1-omni-600M, would help clarify the models’ practical scope. Liquid AI says the smaller model remains under development; the release does not specify a date for a stable version or additional benchmark results.
Key Questions
What did Liquid AI release?
Liquid AI released d1-3B and d1-omni-600M, open-weight models intended to return structured answers to decision tasks in a single forward pass.
What does “single forward pass” mean here?
Liquid AI describes the models as producing a structured decision, such as a classification or score, rather than generating a longer sequence of tokens. The release presents this design for bounded tasks such as request routing or urgency assessment.
How fast is d1-3B on edge devices?
Liquid AI reports one-question response times of 16 milliseconds on a Jetson AGX Thor, 26 milliseconds on a Jetson AGX Orin 64 GB and 50 milliseconds on a Jetson Orin Nano. These are company-reported measurements and may not match other devices or workloads.
Are the models’ multimodal capabilities independently benchmarked?
The release does not provide independent evaluations or published vision and audio benchmark scores. Liquid AI says d1-3B accepts text and images, while d1-omni-600M supports text paired with an image or audio.
Where can developers access the models?
Liquid AI says both sets of weights are available on Hugging Face and points to demos in its System One Arcade Hugging Face Space. Its instructions specify Transformers version 5.14 or later and require loading the models with supplied code enabled.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
