🔍 Read the full analysis: Multimodal Open D1 Decision Models Bring AI Decision-Making To 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 return structured decisions in a single forward pass: d1-3B and the experimental d1-omni-600M. The company reports strong results on selected text benchmarks and response times below 50 milliseconds for d1-3B on tested edge devices, but independent evaluations and published vision and audio benchmark results are absent.
Liquid AI has released d1-3B and d1-omni-600M, open-weight models designed to produce structured decisions in a single forward pass rather than generate a sequence of text, as described in the original analysis of the models. 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 smaller omni model is described as an early research release, and the reported results have not been independently replicated in the material provided.
The models target tasks where an application needs a bounded response, such as classifying a request, assigning it to a team, rating urgency or answering a question about an image. Liquid AI says its decision-model design returns structured answers in a single pass. That differs from using a general-purpose text generator to produce a longer response, though the announcement does not provide comparative testing across real deployments.
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 LFM2.5-Encoder-350M, a bidirectional encoder, with added vision and audio encoders. Liquid AI says it can process text paired with an image or audio, but calls the model experimental and 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. The company’s table lists Decider 4B at 81.1 and Decider 2B at 77.1. Individual results vary: d1-3B scores below Decider 4B on BoolQ, MASSIVE intent and XNLI, so the reported average does not mean it leads on every test.
Testing Decisions on Edge Devices
The release addresses a practical design choice for developers: whether a model can make a narrow decision close to where data is collected, without sending every request to a larger remote system. Low response times and smaller model sizes may matter in devices with limited connectivity, strict latency needs or constrained computing resources. Potential applications include routing support requests or sorting sensor inputs, but the announcement does not establish that either model is ready for any particular production use.
Liquid AI reports that d1-3B answered one question in 16 milliseconds 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. These figures come from the company’s testing, conducted with NVIDIA for the reported GPU and Jetson measurements; actual timing can vary with hardware, software, input and task.
The company also says three questions took 1.3 times as long as one on the tested devices, with the Thor result increasing from 16 to 20 milliseconds. That may be relevant to workloads that can group requests, but the measurement alone does not predict throughput or response times in a deployed application. Liquid AI reports no speed results for d1-omni-600M.
As an affiliate, we earn on qualifying purchases.
Benchmarks Cover Selected Text Tasks
The reported evaluation focuses on seven public datasets: SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI and PAWS-X. They cover a selection of reading, classification, question-answering and language-understanding tasks. The mean scores are useful as a summary of those tests, but they do not measure every kind of decision a developer might build, or establish performance on a specific customer’s data.
Liquid AI says d1-3B retains vision capabilities from its vision-language backbone and that d1-omni-600M handles its supported modalities. However, the release provides no vision or audio benchmark scores. The company says Decision Index version 0.3 includes a private vision split and that audio decision benchmarks remain an open problem. It also describes d1-3B as the “best decision model under 10B on the Decision Index 0.2.1”; that is the company’s characterization of results on the named index, not an independent assessment of broad performance.
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 release instructions specify Transformers version 5.14 or later and loading the models with their supplied code enabled. Access to weights makes testing possible, but does not by itself verify the company’s benchmark or latency claims.
““Best decision model under 10B on the Decision Index 0.2.1.””
— Liquid AI
real-time image and audio processing hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent and Multimodal Results Missing
The source material provides no independent evaluation, confidence intervals or enough methodological detail to determine how closely the benchmark conditions match a particular deployment. The seven-dataset average does not establish accuracy, reliability or safety across all decision tasks. Dataset-level differences also show why the mean should not be read as a result that applies uniformly to each capability.
Vision and audio performance are less documented than the text results: Liquid AI publishes no benchmark scores for either modality, and no speed measurements for d1-omni-600M. The release also does not explain how the models handle ambiguous inputs, how often structured decisions may need human review, or how results change under varied production workloads. Those questions remain open for developers evaluating the models.
As an affiliate, we earn on qualifying purchases.
Developer Testing Will Set the Baseline
The next step is for developers and other evaluators to test the open weights against their own tasks, hardware and operating conditions, and to compare results with suitable alternatives. Independent replication would help establish whether the reported scores and timings hold under other evaluation setups. More detailed vision and audio results would be needed to assess the multimodal claims beyond the supported input formats.
Liquid AI identifies d1-omni-600M as a research release, so further development may bring changes to its behavior or documentation. The announcement does not provide a timetable for additional evaluations or updates. Until more results are available, the company’s figures should be treated as reported measurements rather than a guarantee of performance in a specific application.
AI decision models for embedded systems
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 the omni model as experimental and under development.
What tasks are the models intended to handle?
Liquid AI presents them for decision tasks such as classifying or routing requests, rating urgency and answering questions about images. The release does not establish performance across all possible applications.
How fast is d1-3B on edge hardware?
Liquid AI reports a response time of 16 milliseconds on Jetson AGX Thor, 26 milliseconds on Jetson AGX Orin 64 GB and 50 milliseconds on Jetson Orin Nano for one question. These are company-reported measurements, and results may vary by task and setup.
Are the benchmark results independently verified?
Not in the source material. The scores and timing figures are reported by Liquid AI; independent replication is not provided.
Are vision and audio performance results available?
The release does not provide vision or audio benchmark scores. It describes the models’ supported inputs, but more testing is needed to assess performance on those modalities.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
