🔍 Read the full analysis: Building Edge AI Systems With Multimodal Open D1 Decision Models 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 has 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 d1-3B responses 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 to decision tasks in a single forward pass. The company says d1-3B scored 48.57 on Decision Index 0.2.1 and answered a question in 16 milliseconds on an NVIDIA Jetson AGX Thor; the release does not include independent replication of those results.
The models target tasks where a system needs to classify, score or route an input rather than generate a long response. Liquid AI describes examples such as routing customer requests, judging urgency and answering questions about images. Both models are based on the company’s Liquid Foundation Models, but use different foundations and support different inputs.
d1-3B is built on the 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. It supports text paired with an image or text paired with audio. Liquid AI describes it as an early research release that remains under development.
Across seven public datasets, Liquid AI reports mean scores of 82.9 for d1-3B and 78.4 for d1-omni-600M. The company’s comparison table lists scores of 81.1 for Decider 4B and 77.1 for Decider 2B. Results vary by dataset: d1-3B scores below Decider 4B on BoolQ, MASSIVE intent classification and XNLI. These figures reflect the company’s selected evaluation set, not a general measure of performance across all decision applications.
Why Edge Decision Speed Matters
For developers deploying models on devices near where data is collected, response time and hardware requirements can shape whether a system is practical. Liquid AI reports that d1-3B answered a question in 16 milliseconds on Jetson AGX Thor, 26 milliseconds on Jetson AGX Orin 64 GB and 50 milliseconds on Jetson Orin Nano. The company says responses took less than 50 milliseconds on the tested devices, though the reported Nano result is exactly 50 milliseconds.
The release also reports that three questions took 1.3 times as long as one on tested devices; on the AGX Thor, the measured time rose from 16 to 20 milliseconds. That result may be relevant to workloads that can group requests, but it does not establish performance in a particular product. Device configuration, input and task can affect real-world latency.
A model designed to produce a predefined decision could suit applications that do not need a full generative response. The smaller omni model may interest teams with tight hardware constraints, but its reported mean on selected datasets does not establish that it will be more accurate or dependable for a particular deployment.
NVIDIA Jetson AGX Orin edge AI device
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Benchmarks Cover
Liquid AI reports evaluation on seven public datasets covering reading comprehension, toxicity detection, intent classification, medical question answering and cross-lingual understanding. The named datasets include SQuAD 2.0, Civil Comments and MASSIVE, as well as PubMedQA, BoolQ, XNLI and PAWS-X. A mean score summarizes performance over that specific selection; individual results differ, and the release’s comparisons do not settle how the models perform on other tasks.
For speed testing, Liquid AI says it worked with NVIDIA and measured d1-3B on NVIDIA GPUs and Jetson devices, alongside Apple M5 Pro and AMD MI325X systems. It reports 8 milliseconds per question on an NVIDIA RTX 4090 and 9 milliseconds on an AMD MI325X. These are company-reported measurements. The release provides no speed results for d1-omni-600M, which it identifies as an experimental model.
Liquid AI says d1-3B retains vision capabilities from its vision-language foundation and that d1-omni-600M handles its supported modalities. However, the announcement supplies no vision or audio benchmark scores. The company also says Decision Index version 0.3 has only a private vision split and that audio decision benchmarks remain an open problem.
“Best decision model under 10B on the Decision Index 0.2.1”
— Liquid AI
vision-language AI models for decision making
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Release Does Not Establish
The reported results have not been independently replicated in the supplied material, and the release does not provide confidence intervals or enough evaluation detail to determine how closely the tests match specific production settings. The seven datasets measure selected capabilities; they do not establish accuracy, reliability or safety across all decision tasks.
Real-world handling of ambiguous inputs, the frequency of human review and performance under varied workloads are also not specified. Vision and audio performance remains especially unclear: the release offers no modality-specific benchmark scores, and it provides no speed measurements for d1-omni-600M. Liquid AI calls that model an early research release, so its capabilities and operating characteristics may change as development continues.
audio and image processing AI hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing the Models in Deployment
Both models are available as open weights on Hugging Face, and Liquid AI points users to demonstrations in its System One Arcade Hugging Face Space. The company’s instructions specify Transformers version 5.14 or later and require users to load the models with their supplied code enabled.
The next practical test is how the models perform on developers’ own inputs, devices and decision criteria. Independent evaluations, published vision and audio results, and additional speed data—particularly for d1-omni-600M—would help clarify how far the company’s reported benchmark and latency results extend beyond the release’s stated tests.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Liquid AI release?
It released d1-3B and d1-omni-600M, open-weight models intended to classify, score or otherwise answer decision tasks with a structured output in one forward pass.
What inputs can the models process?
d1-3B accepts text and images. Liquid AI says d1-omni-600M supports text paired with an image or text paired with audio; the company describes it as an early research model still under development.
How fast is d1-3B on edge hardware?
Liquid AI reports one-question response 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 measurements, not independently replicated results in the supplied material.
Are the benchmark claims independently verified?
The supplied release reports Liquid AI’s results on seven public datasets, but does not provide independent evaluations. Performance on other tasks and in production remains unestablished.
Where can developers access the models?
Liquid AI says both models are available as open weights on Hugging Face and links to demos in its System One Arcade Hugging Face Space. Its instructions call for Transformers 5.14 or later and loading with the supplied code enabled.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
