TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Liquid AI released two open-weight models in its d1 decision family: d1-3B for text and image inputs, and experimental d1-omni-600M for text with either images or audio. The company reports strong benchmark results and sub-50-millisecond single-question inference for d1-3B on tested edge devices; the audio-capable model has no published speed results.
Liquid AI has released two open-weight decision models intended for local and edge use: d1-3B, which accepts text and images, and the experimental d1-omni-600M, which accepts text with either images or audio. The release matters for developers building systems that need to classify, route or score inputs quickly without relying on a model generating a long sequence of tokens, much like other decision models.
Liquid AI says its decision models produce answers in a single forward pass, rather than generating output token by token like a conventional text-generation model. The company describes d1-3B as its higher-capacity model and d1-omni-600M as a smaller option, joining a growing family of compact decision models. Both are based on the company’s Liquid Foundation Models, but use different underlying architectures: d1-3B builds on the decoder-only LFM2.5-VL-3B, while d1-omni-600M builds on the bidirectional LFM2.5-Encoder-350M with added vision and audio encoders.
On seven public datasets covering reading comprehension, toxicity detection, intent classification, medical question answering and cross-lingual understanding, d1-3B recorded a mean score of 82.9, compared with 81.1 for Decider 4B in Liquid AI’s table. The 600-million-parameter d1-omni-600M scored 78.4, above the listed 77.1 for Decider 2B. These are results reported by Liquid AI; the release does not include independent validation of the comparison.
For d1-3B, Liquid AI reports a single-question response time of 16 milliseconds on a Jetson AGX Thor, 26 milliseconds on a Jetson AGX Orin 64 GB and 50 milliseconds on a Jetson Orin Nano. The company says it measured the model with NVIDIA inferences on edge devices and desktop or data-center GPUs. It reports no speed figures for d1-omni-600M because that model remains an early research release.
Fast Decisions on Edge Hardware
The release targets applications where a system needs to act on incoming information rather than write a conversational response. Examples in Liquid AI’s demonstration include identifying refund requests, assigning customer support tickets to teams, rating urgency and answering questions about images. If the reported performance holds in a developer’s deployment, these tasks could be handled on nearby hardware, potentially reducing the need to send every input to a remote service.
Latency and model size are central to the pitch. A single-pass decision model may suit repeated classification or routing tasks, while local inference can be relevant where connectivity, response time or data handling requirements favor on-device processing. The company’s figures indicate that d1-3B can run on several NVIDIA Jetson devices, including the Orin Nano, but they do not establish performance for every application or hardware configuration.
The modality options also broaden the possible inputs. d1-3B works with text and images; d1-omni-600M can take text with images or text with audio. However, Liquid AI has not published vision or audio benchmark scores for this release, so its claims about those capabilities have less public comparative evidence than the text-focused dataset results.
As an affiliate, we earn on qualifying purchases.
How the Two Models Differ
Liquid AI presents d1 as a separate model family from its generative models. Rather than producing a sequence of tokens, a d1 model receives a state and specified questions, then returns structured outputs such as a choice or score. The company’s example processes several named decisions about one customer message in a single call, and also demonstrates image input and batching multiple requests.
The two releases differ in scale and maturity. d1-3B has 3 billion parameters and inherits its text-and-image design from LFM2.5-VL-3B. The smaller d1-omni-600M has 600 million parameters and uses an encoder backbone with added vision and audio encoders. Liquid AI calls the latter experimental and says it is still undergoing development.
The results in the release cover seven public datasets, but not vision or audio decision tests. Liquid AI says the Decision Index version 0.3 includes a private vision split and that audio decision benchmarks remain an open problem. It also reports d1-3B as the top model under 10 billion parameters on Decision Index 0.2.1, with a score of 48.57, ahead of the models it compared, including Decider 35B-A3B at 47.11. These rankings reflect the company’s reported evaluation and the named index version.
“Unlike our generative models, decision models don’t produce tokens but answer in a single forward pass.”
— Liquid AI
As an affiliate, we earn on qualifying purchases.
Open Questions on Audio and Performance
Independent replication is not included in the supplied release, and the benchmark and speed figures are Liquid AI’s own reported results. The source does not provide enough detail to establish how results would change across different software versions, workloads, power limits or deployment settings. Its device timings should be read as measurements under the company’s test conditions, not a guarantee for every use.
The release also leaves important modality questions unresolved. Liquid AI says it has not reported vision or audio benchmark scores, and it publishes no speed results for d1-omni-600M. That means readers cannot use the release to compare the smaller model’s latency or decision quality on audio tasks against other systems. The exact publication date is not stated in the source material, which identifies the release as a 2026 announcement.
As an affiliate, we earn on qualifying purchases.
Testing the Models in Practice
Both models are available as open weights on Hugging Face, and Liquid AI points developers to demonstrations in its System One Arcade. The company’s setup instructions require Transformers 5.14 or later and note that the models ship with their own code, which must be loaded with the relevant remote-code option. Developers evaluating the models will need to test them against their own decision tasks and hardware rather than relying on headline benchmark scores alone.
The next useful evidence would include broader independent evaluations, published tests for image and audio decisions, and performance measurements for d1-omni-600M. Liquid AI says that model is still under development; the release does not provide a timetable for a stable version or for additional benchmark results.
on-device image and audio classifier
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 two open-weight decision models: d1-3B, which accepts text and images, and experimental d1-omni-600M, which accepts text with either images or audio.
How are decision models different from text generators?
Liquid AI says d1 models return answers in a single forward pass, rather than generating a response one token at a time. The intended outputs include structured decisions such as classifications, choices and scores.
How fast is d1-3B on edge devices?
Liquid AI reports single-question times of 16 milliseconds on Jetson AGX Thor, 26 milliseconds on Jetson AGX Orin 64 GB and 50 milliseconds on Jetson Orin Nano. Those are company-reported measurements under its test conditions.
Does d1-omni-600M have published audio benchmarks?
No. Liquid AI says it has not reported audio or vision benchmark scores for this release. It also does not publish speed numbers for d1-omni-600M, which it describes as an early research model.
Where can developers get the models?
Liquid AI says both open-weight models are available on Hugging Face. Its release also links to demos in the System One Arcade and provides setup guidance for d1-3B.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
