🔍 Read the full analysis: What Makes Multimodal Open D1 Decision Models Suited To Edge AI? 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 two open-weight models designed to return structured decisions in a single forward pass: d1-3B and experimental d1-omni-600M. The company reports sub-50-millisecond d1-3B responses on several tested edge devices and benchmark scores across seven public datasets, but independent evaluations and published vision or audio benchmark results are not included.
Liquid AI has released d1-3B and d1-omni-600M, two open-weight models designed to produce structured decisions in a single forward pass rather than generate a sequence of text tokens. The company says d1-3B answered a test question in 16 milliseconds on an NVIDIA Jetson AGX Thor; its results could interest developers building decision tools for devices where latency or computing capacity matters, though the measurements have not been independently replicated in the material provided.
Liquid AI says the models are built on its Liquid Foundation Models and target tasks such as classifying requests, scoring urgency and answering questions about images. The intended output is a structured answer for a defined decision task, not a free-form conversational response. That design may suit applications that need a quick classification or routing result, but the announcement does not establish how well the models handle every production scenario.
d1-3B accepts text and images and is based on the company’s LFM2.5-VL-3B vision-language model. The smaller d1-omni-600M uses the LFM2.5-Encoder-350M bidirectional encoder with added vision and audio encoders; Liquid AI says it can process text paired with an image or audio. The company describes this model as an early research release that is still under development.
Across 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 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 and XNLI. The reported means are company evaluations, not independent confirmation.
Latency Could Suit Local Decisions
For edge deployments, a model’s response time and hardware requirements can determine whether it is practical to run near the source of data. Liquid AI’s reported tests put d1-3B at 26 milliseconds on Jetson AGX Orin 64 GB and 50 milliseconds on Jetson Orin Nano, in addition to the 16-millisecond result on Jetson AGX Thor. The company says a single question took less than 50 milliseconds on the tested Jetson devices.
Keeping a decision model on a device could be useful in settings where a fast response, limited connectivity or local processing is preferred. Those are potential applications, not outcomes established by the release. Actual speed and usefulness will depend on the hardware, software configuration, inputs and task. The published tests do not show that every edge device or deployed application will meet the same response times.
Liquid AI also reports that three questions took 1.3 times as long as one on tested devices; on AGX Thor, the reported time rose from 16 to 20 milliseconds. This may indicate that grouped questions are worth testing for some workloads, but the figures do not establish production throughput or performance under sustained use. The smaller model’s reported mean above the listed Decider 2B score is another point for developers to evaluate, not proof of broader superiority.
NVIDIA Jetson AGX Orin developer kit
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Benchmarks Measure
The company’s comparison covers seven public text-focused datasets: SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI and PAWS-X. These datasets examine selected capabilities, and an average across them cannot show how a model will perform on every decision task. Liquid AI reports the highest mean in its table for d1-3B, while d1-omni-600M’s mean is above the listed Decider 2B score; individual results are not uniformly higher.
For speed testing conducted with NVIDIA, Liquid AI reports measurements on NVIDIA GPUs and Jetson devices, as well as Apple M5 Pro and AMD MI325X hardware. It lists 8 milliseconds per question on an NVIDIA RTX 4090 and 9 milliseconds on an AMD MI325X. The release provides no speed measurements for d1-omni-600M, which the company identifies as experimental. 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 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.
Independent and Multimodal Results
The release does not include independent benchmark replication, confidence intervals or enough testing detail to determine how closely the reported setup matches a particular deployment. The seven datasets do not establish accuracy, reliability or safety across all decision tasks. It is also not clear how the models handle ambiguous inputs, how often a person should review their decisions, or how they perform under varied production workloads.
Liquid AI says it checked whether d1-3B retained vision capabilities from its underlying model and whether d1-omni-600M handled its supported modalities, but it provides no vision or audio benchmark scores in the release. The company says Decision Index version 0.3 includes only a private vision split and that audio decision benchmarks remain an open problem. No speed results are given for d1-omni-600M. Those gaps limit what can be concluded about the models’ multimodal performance, despite their stated input support.
As an affiliate, we earn on qualifying purchases.
Testing on Target Devices
The models’ open weights let developers evaluate them against their own tasks and hardware. That testing will be needed to establish whether reported latency and dataset scores translate to a particular application, including whether outputs require human review. Liquid AI’s release instructions specify Transformers 5.14 or later and require loading the models with the supplied code enabled.
Further evidence would include independent evaluations, clearer testing details, production workload measurements and published results for image and audio decision tasks. Liquid AI describes d1-omni-600M as still under development, but the source material does not give a date for additional releases or benchmarks. Until such information is available, the reported figures should be treated as company measurements rather than a guarantee of deployment performance.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Liquid AI release?
Liquid AI released d1-3B and d1-omni-600M as open-weight models intended to return structured decisions in a single forward pass. The company describes d1-omni-600M as an experimental research release.
What does “single forward pass” mean here?
The models are designed to return a structured answer for a decision task, such as a classification or score, rather than generate a sequence of tokens as a text-generating model does. The release does not establish that this approach is best for every task.
How fast was 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-reported test results, not independently replicated figures.
Are the models’ image and audio capabilities independently benchmarked?
No such results are provided in the release. Liquid AI says d1-3B accepts text and images and d1-omni-600M supports text paired with an image or audio, but it does not publish vision or audio benchmark scores.
Where can developers access the models?
Liquid AI says both models are available as open weights on Hugging Face, with demos in its System One Arcade Hugging Face Space. The release instructions call for Transformers 5.14 or later and use of the supplied loading code.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
