🔍 Read the full analysis: Multimodal Open D1 Decision Models For The Edge 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 has released d1-3B and experimental d1-omni-600M, open-weight models designed to return structured decisions in a single forward pass. The company reports selected benchmark scores and sub-50-millisecond d1-3B responses on tested devices, but independent evaluations and published vision or audio benchmark results are absent.
The original analysis describes Liquid AI’s d1 approach. Liquid AI has released d1-3B and d1-omni-600M, two open-weight models designed to return structured answers for tasks such as classification and routing in a single forward pass. The company says d1-3B responded in 16 milliseconds on an NVIDIA Jetson AGX Thor and reports benchmark scores for both models, positioning them for decision workloads on edge devices; independent performance results were not provided.
The models are built on Liquid AI’s Liquid Foundation Models and target decisions that can be expressed as structured outputs rather than open-ended text generation, amid the wider development of open models. Examples in the release include assigning a customer request to a team, estimating urgency, and answering a question about an image. The company describes the approach as a fit for deployments where latency and device constraints matter, a focus shared by efforts to shape AI collaboration.
d1-3B is based on the LFM2.5-VL-3B vision-language model and accepts text and images. The smaller d1-omni-600M is based on the LFM2.5-Encoder-350M bidirectional encoder, with vision and audio encoders added. It supports text paired with an image or audio, and Liquid AI calls it 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. Its comparison table lists Decider 4B at 81.1 and Decider 2B at 77.1. The company says d1-3B scored 48.57 on Decision Index 0.2.1. These are company-reported results, and scores vary by dataset: d1-3B is below Decider 4B on BoolQ, MASSIVE intent and XNLI.
Liquid AI 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 reports 8 milliseconds per question on an NVIDIA RTX 4090 and 9 milliseconds on an AMD MI325X. The release says three questions took 1.3 times as long as one on tested devices, with the AGX Thor measurement rising from 16 to 20 milliseconds. The release does not provide speed figures for d1-omni-600M.
Decision Models on Edge Hardware
For developers, a model that produces a predefined decision rather than a long generated response could suit products that need to sort, score or route inputs quickly. The reported edge-device timings offer an initial indication of how d1-3B performed on selected hardware, while the open weights allow teams to test it against their own tasks and constraints.
The smaller omni model may also interest teams evaluating multimodal decisions on limited hardware. Liquid AI’s reported mean score for d1-omni-600M is higher than the listed Decider 2B result on its selected datasets, despite having one-quarter as many parameters, according to the company. That comparison is limited to this evaluation: it does not establish performance across other applications, or demonstrate the model’s vision and audio quality.
The practical value will depend on more than a benchmark average or a single response-time measurement. Developers still need to check accuracy, reliability and review requirements on their own inputs and devices. The release provides a reason to test these models, not evidence that they will outperform alternatives in every deployment.
As an affiliate, we earn on qualifying purchases.
What the Evaluations Cover
Liquid AI evaluated the models on seven public datasets: SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI and PAWS-X. The categories span reading comprehension, toxicity detection, intent recognition, medical question answering and cross-lingual understanding. The company’s reported mean combines results across this selected set; individual results are not uniformly higher than the listed comparison models.
The release says d1-3B retained vision capabilities from its LFM2.5-VL-3B base, and says d1-omni-600M handles its supported modalities. However, it supplies no vision or audio benchmark scores. Liquid AI says Decision Index version 0.3 has only a private vision split and that audio decision benchmarks remain an open problem. Its speed tests were conducted with NVIDIA and included NVIDIA GPUs and Jetson devices, as well as Apple M5 Pro and AMD MI325X; the company does not report independent replication.
Both models are available as open weights on Hugging Face, and Liquid AI points to demos in its System One Arcade Hugging Face Space. Its release instructions call for Transformers version 5.14 or later and loading the models with their supplied code enabled.
““Best decision model under 10B on the Decision Index 0.2.1””
— Liquid AI
As an affiliate, we earn on qualifying purchases.
Limits of the Reported Results
The announcement does not include independent evaluations, confidence intervals or enough information to determine how closely its benchmark setup matches a particular deployment. Seven public datasets measure selected capabilities; they do not establish accuracy, safety or reliability across all decision tasks. The reported speed figures may also vary with device configuration, input, software and workload.
Performance on vision and audio remains undocumented in published benchmark scores, and no latency results are given 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 they perform under varied production conditions. Those questions remain for further testing.
audio and image processing devices
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing the Models in Deployment
The immediate next step for interested developers is to download the open weights and evaluate them on relevant tasks and hardware, following Liquid AI’s stated software and loading requirements. Such tests can show whether the reported latency and dataset scores translate to a particular use case, while checking how the model handles uncertain or difficult inputs.
Further evidence would include independent replications, detailed vision and audio evaluations, and measurements for d1-omni-600M. Liquid AI describes that model as under development, but the announcement does not give a schedule for updates or additional results. Until those details are available, the reported figures should be treated as company measurements on specified tests, not general performance guarantees.
low latency AI inference edge device
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 intended to produce structured decisions in a single forward pass.
What kinds of inputs do the models support?
d1-3B accepts text and images. Liquid AI says d1-omni-600M can process text with an image or text with audio; the company describes it as an experimental model still in development.
Are the reported benchmark and speed results independently verified?
The source material provides Liquid AI-reported results and says independent replication was not included. The release also does not publish confidence intervals.
Where can developers access the models?
Liquid AI says both models are available as open weights on Hugging Face and points users 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
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
