🔍 Read the full analysis: Multimodal AI Decisions, Powered By Open D1 Models At The Edge 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 produce structured answers for decision tasks in a single forward pass. The company reports benchmark scores and sub-50-millisecond response times for d1-3B on tested devices, but the release does not include independent evaluations or vision and audio benchmark results.
Liquid AI has released d1-3B and d1-omni-600M, open-weight models intended to classify, score and answer certain questions with structured outputs in a single forward pass, as described in the original analysis. The company says d1-3B returned an answer in 16 milliseconds on an NVIDIA Jetson AGX Thor, positioning the models for uses where responses need to run close to the source of data; the published performance figures are company-reported, not independently replicated.
The models are built on Liquid AI’s Liquid Foundation Models and are designed for tasks such as routing a customer request, judging urgency or answering questions about an image, amid wider industry moves toward open AI models. Unlike a conventional generative approach that produces a sequence of tokens, the company describes these models as returning structured answers for predefined decision tasks. Their intended role is narrower than general-purpose text generation.
d1-3B is based on the company’s LFM2.5-VL-3B vision-language model and accepts text and images. d1-omni-600M uses the LFM2.5-Encoder-350M bidirectional encoder with added vision and audio encoders; it can process text paired with an image or text paired with audio. Liquid AI labels the smaller model 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, in a year of shifting open-model trends. Its comparison table lists scores of 81.1 for Decider 4B and 77.1 for Decider 2B. Results varied by dataset: d1-3B scored below Decider 4B on BoolQ, MASSIVE intent and XNLI, so the reported averages do not show it leading across every test.
Edge Decisions Without Full Generation
The release offers developers another option for applications that need a bounded classification or routing decision rather than a long generated response. If a model can perform an appropriate task on a nearby device, it may fit products where response time, network access or device resources shape the design. Liquid AI’s tests give an initial indication of speed on selected hardware, not a guarantee for a deployed system.
The smaller model’s reported dataset average could also interest teams with limited compute budgets. But a parameter count and an average across seven datasets do not establish that it will outperform larger systems on a particular product’s inputs. For consequential decisions, teams would still need to test accuracy, error patterns, and whether human review is required. Those questions are especially relevant because the release provides no independent assessment of safety or reliability across real-world workflows.
NVIDIA Jetson AGX Thor AI development kit
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Published Tests Cover
The seven-dataset evaluation includes SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI and PAWS-X. These public tests cover selected language capabilities; they are not a broad assessment of every decision task or operating environment. Liquid AI says it also checked whether d1-3B retained vision capabilities from its underlying model and whether d1-omni-600M handled supported modalities, but the release gives no vision or audio benchmark scores.
The company reports that d1-3B took 16 milliseconds for one question on Jetson AGX Thor, 26 milliseconds on Jetson AGX Orin 64 GB and 50 milliseconds on Jetson Orin Nano. It also reports 8 milliseconds per question on an NVIDIA RTX 4090 and 9 milliseconds on an AMD MI325X. Liquid AI says the timing tests were conducted with NVIDIA for its NVIDIA hardware measurements. The figures describe the company’s test setup; the announcement does not provide independent replication or establish performance for other inputs and software configurations.
Liquid AI also reports that three questions took 1.3 times as long as one on tested devices, with the AGX Thor result rising from 16 to 20 milliseconds. This is a reported result for those tests, not evidence that batching will have the same effect in every workload. The announcement says d1-omni-600M has no speed results because it remains an early research model.
“Best decision model under 10B on the Decision Index 0.2.1”
— Liquid AI
As an affiliate, we earn on qualifying purchases.
Independent and Multimodal Evidence
The release does not provide independent evaluations, confidence intervals or enough methodological detail to establish how the benchmark setup compares with a developer’s intended use. The scores reflect seven selected public datasets; they do not establish accuracy, reliability or safety across all decision tasks. The company’s reported results should be read within that scope.
Evidence is thinner for image and audio use. Liquid AI publishes no vision or audio benchmark scores, and says Decision Index version 0.3 includes only a private vision split while audio decision benchmarks remain an open problem. The announcement also does not explain how the models handle ambiguous inputs, how often structured answers need human review, or how results change under varied production workloads. Those details remain unanswered in the release.
Because d1-omni-600M is described as experimental and under development, its capabilities and operational characteristics may change. The company has made the weights available, but availability alone does not establish that the model is ready for a particular deployment.
As an affiliate, we earn on qualifying purchases.
Testing the Models in Practice
Liquid AI has made both models 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 say to load the models with their supplied code enabled. The announcement does not give a schedule for additional benchmark results or a development milestone for the experimental omni model.
The next useful evidence will come from evaluations that disclose their methods and test the models on task-specific inputs and hardware. Developers can compare results on their own workloads, including error rates and review requirements, while treating Liquid AI’s timing and score figures as company-reported measurements until independently checked.
audio and image processing AI tools
As an affiliate, we earn on qualifying purchases.
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 return structured answers for 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 does not establish that this design is suitable for every decision task.
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-reported test results, not independent measurements.
Are the benchmark results independently verified?
The source material does not provide independent evaluations. The reported mean scores come from Liquid AI’s tests on seven public datasets, and results varied across individual datasets.
Can the models handle images and audio?
d1-3B accepts text and images. Liquid AI says d1-omni-600M can process text with an image or text with audio, but the release provides no vision or audio benchmark scores; the omni model is also described as experimental.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
