🔍 Read the full analysis: From Multiple Inputs To Edge Decisions: Exploring Open D1 Models 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, d1-3B and d1-omni-600M, designed to produce structured answers for decision tasks in a single forward pass. The company reports benchmark scores and sub-50-millisecond responses for d1-3B on tested edge devices, but independent results and vision and audio benchmark data are not included in the release.
Liquid AI has released d1-3B and d1-omni-600M, two open-weight models designed to return structured decisions in a single forward pass rather than generate a sequence of tokens, as described in the original analysis. The company says d1-3B reached 48.57 on Decision Index 0.2.1 and answered a question in 16 milliseconds on an NVIDIA Jetson AGX Thor; the published evaluations and speed measurements are company-reported, with no independent replication supplied in the release.
The models are intended for tasks that can be framed as a decision or structured answer, such as routing a customer request, assessing urgency, or answering a question about an image. Liquid AI says this design may suit applications where latency, hardware capacity or running inference near the data source matters. It is distinct from using a general-purpose text generator to produce a longer response, although the announcement does not establish how the models perform in particular deployed products.
d1-3B is based on Liquid AI’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 with audio. Liquid AI describes the omni model as an early research release that remains 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 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. These comparisons reflect the company’s selected evaluation set, not performance across all decision tasks.
Edge Decisions on Smaller Devices
The release targets developers who need a model to make a bounded decision without relying on a larger text-generation system. In settings such as a device sorting incoming requests or interpreting information close to where it was collected, response time and hardware requirements can affect whether a model is practical. A smaller model that returns a structured result may be useful for those workflows, but suitability depends on the task and on how errors are handled.
Liquid AI reports that d1-3B took 16 milliseconds per question on a Jetson AGX Thor, 26 milliseconds on a Jetson AGX Orin 64 GB and 50 milliseconds on a Jetson Orin Nano. The company also says three questions took 1.3 times as long as one on tested devices; on the AGX Thor, its reported time rose from 16 to 20 milliseconds. Those measurements offer an initial indication of latency under the company’s test setup, not a guarantee for a different device, software configuration or production workload.
The reported mean for d1-omni-600M is above the listed Decider 2B result despite the smaller model’s lower parameter count. That may interest teams with tight compute limits, but a mean across seven datasets does not show how often the model will make a correct or safe decision in a specific application. The release does not provide independent testing or enough deployment evidence to settle that question.
As an affiliate, we earn on qualifying purchases.
What the Benchmarks Cover
Liquid AI’s reported evaluation uses seven public datasets: SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI and PAWS-X. They cover selected language tasks, including reading comprehension, intent recognition, toxicity detection, medical question answering and cross-lingual understanding. The aggregate scores summarize that particular selection; individual results differ, and the set does not represent every kind of decision a developer might want a model to make.
The release says d1-3B retained vision capabilities from its vision-language backbone and that d1-omni-600M handles its supported modalities. However, Liquid AI provides no vision or audio benchmark scores in the material described here. It says Decision Index version 0.3 has only a private vision split and that audio decision benchmarks remain an open problem. Accordingly, the published public-dataset averages should not be treated as evidence of equivalent performance on image or audio tasks.
For speed, 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, and it reports no speed results for d1-omni-600M because that model is still experimental.
“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 Published Results
The announcement does not include independent evaluations or confidence intervals, and the information provided does not make clear how closely the benchmark setup matches a particular deployment. The seven public datasets test selected capabilities; their scores do not establish reliability, accuracy or safety across all decision tasks. The company’s averages are best understood as results on that limited selection.
Performance on images and audio is especially hard to judge from the release because Liquid AI supplies no modality-specific benchmark scores, and it provides no speed measurements for d1-omni-600M. The company also identifies that model as experimental, so its capabilities and operational characteristics may change as development continues.
Other deployment questions remain unanswered: how the models respond to ambiguous inputs, how often their structured outputs require human review, and how performance changes under varied production workloads. The availability of open weights allows developers to test the models, but does not by itself answer those questions or establish suitability for a particular use.
As an affiliate, we earn on qualifying purchases.
Testing Models on Local Workloads
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 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 compare model outputs, latency and error rates against their own tasks and hardware, rather than assume the published averages will transfer directly.
Further independent evaluations, more detailed deployment testing, and vision and audio benchmarks would help clarify how broadly the results apply. Liquid AI has not provided a date for such results in the supplied material. For now, d1-3B has reported text-and-image capabilities and hardware timings, while d1-omni-600M remains an early research model with no published speed figures.
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 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 or answer in one forward pass, rather than generating a sequence of tokens. The announcement positions them for bounded tasks such as routing requests or assessing urgency.
How fast is d1-3B on edge hardware?
Liquid AI reports one-question 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 and may not match results on other setups.
Have the benchmark results been independently verified?
The supplied release material does not include independent replication or confidence intervals. The scores and timing figures should be treated as company-reported results.
Are vision and audio capabilities benchmarked?
Liquid AI says the models support specified image and audio inputs, but it does not provide vision or audio benchmark scores in the release. It also publishes no speed figures for the experimental d1-omni-600M model.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
