🔍 Read the full analysis: Why Use Multimodal Open D1 Decision Models For Edge Applications? 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 d1-3B and experimental d1-omni-600M, open-weight models designed to return structured decisions in a single forward pass. The company reports benchmark and speed results for d1-3B, including a 16-millisecond response on an NVIDIA Jetson AGX Thor; independent replication and published vision and audio scores are not provided.
As detailed in the original analysis, Liquid AI has released d1-3B and d1-omni-600M, open-weight models designed to classify, score and answer decision tasks in a single forward pass. The company says d1-3B returned one answer in 16 milliseconds on an NVIDIA Jetson AGX Thor; the figures are company-reported, and independent evaluations have not been provided.
The models are intended to produce structured answers rather than generate a sequence of tokens as a conventional text-generation system does, reflecting broader evolution in open AI models. Liquid AI describes potential uses including routing customer requests, judging urgency and answering questions about images. The company positions the models for settings where response time and device constraints matter, including edge applications that process data near where it is collected.
d1-3B is based on the company’s 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 audio. Liquid AI identifies the omni model as an early research release that is still under development.
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, amid a rapid cadence of open-model releases. Its comparison table lists Decider 4B at 81.1 and Decider 2B at 77.1. Results vary by dataset: d1-3B scores below Decider 4B on BoolQ, MASSIVE intent and XNLI. The mean scores describe that selected test set, not performance on every decision task.
Why Edge Decision Speed Matters
For developers building systems on cameras, robots, kiosks or other devices, a model’s response time and hardware requirements can affect whether decisions can be made locally. Liquid AI’s reported results give developers a starting point for testing small-footprint decision models without assuming that every task needs a general-purpose generative system. Local processing may also reduce reliance on sending each input to a remote service, though the release does not quantify any cost, privacy or reliability benefits.
The company reports d1-3B response times of 16 milliseconds on Jetson AGX Thor, 26 milliseconds on Jetson AGX Orin 64 GB and 50 milliseconds on Jetson Orin Nano. It also says three questions took 1.3 times as long as one across tested devices; on the AGX Thor, the reported time rose from 16 to 20 milliseconds. These measurements could interest teams evaluating grouped requests, but they are not guarantees for a particular application. Results can vary with hardware, software, input and task.
The smaller model’s reported mean score is also relevant to constrained deployments: Liquid AI says d1-omni-600M exceeded Decider 2B on the selected dataset mean despite having a quarter of its parameters. That comparison is limited to the company’s evaluation and does not establish that the smaller model will perform better across other workloads.
NVIDIA Jetson AGX Xavier AI development kit
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Models Built for Structured Decisions
Liquid AI says the models build on its Liquid Foundation Models and are designed to return a decision in one forward pass. The approach targets bounded tasks—such as assigning an intent or assessing urgency—rather than open-ended text generation. The models’ open weights allow developers to evaluate them in their own environments, but open availability alone does not establish their suitability for production use.
The reported evaluation covers seven public datasets: SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI and PAWS-X. Liquid AI also says it checked that d1-3B retained vision capabilities from its vision-language backbone and that d1-omni-600M handled its supported modalities. The release supplies no vision or audio benchmark scores. The company says Decision Index version 0.3 has only a private vision split and that audio decision benchmarks remain an open problem.
For speed testing, Liquid AI says it worked with NVIDIA to measure d1-3B on NVIDIA GPUs and Jetson devices, and also lists measurements for Apple M5 Pro and AMD MI325X. It reports 8 milliseconds per question on an NVIDIA RTX 4090 and 9 milliseconds on an AMD MI325X. The release does not provide independent replication. It reports no speed measurements for d1-omni-600M.
“Best decision model under 10B on the Decision Index 0.2.1”
— Liquid AI
Evidence Beyond the Published Tests
The reported scores and response times come from Liquid AI’s evaluation; the release does not include independent replication or confidence intervals. The seven datasets cover selected capabilities and do not establish accuracy, reliability or safety across all decision tasks. The benchmark setup may also differ from the input types and workloads developers encounter in deployment.
Performance for vision and audio is less documented: the release provides no scores for either modality and no speed figures for the omni model. It also does not say how the models handle ambiguous inputs, how often decisions require human review, or how their performance changes under varied production workloads. Because d1-omni-600M is experimental, its capabilities and operating characteristics may change as development continues.
Developer Testing and Model Access
Both models are available as open weights on Hugging Face, and Liquid AI points users to demos in its System One Arcade Hugging Face Space. The company’s release instructions specify Transformers version 5.14 or later and require loading the models with the supplied code enabled.
The next practical step for prospective users is to test the models against their own tasks, devices and input conditions, rather than treating reported benchmarks as deployment results. Further independent evaluations, published vision and audio measurements, and performance data for d1-omni-600M would help clarify how well the models transfer beyond the reported tests. The source material does not specify a timetable for those 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 multimodal mean for these models?
d1-3B accepts text and images. Liquid AI says d1-omni-600M can process text paired with an image or audio, though the company labels that model an early research release.
How fast is d1-3B on edge devices?
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 independent results or guarantees for other workloads.
Have the benchmark results been independently verified?
The source material provides company-reported scores across seven public datasets but no independent replication or confidence intervals. The reported averages apply to that selected dataset group.
Where can developers access the models?
Liquid AI says both models are available as open weights on Hugging Face and provides demos through its System One Arcade Hugging Face Space. Its 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.
