Liquid AI Releases Downloadable Models That Make Decisions Without Writing Answers

Developers can now download Liquid AI models that choose an answer without writing one. On October 7, 2026, the company released open-weight d1-3B and experimental d1-omni-600M on Hugging Face. Both make structured decisions in one processing pass, targeting tasks where software needs a choice or score rather than generated text.

Two days from hosted access to open weights

Liquid AI introduced its hosted d1 model on October 5, with text and image support through the company’s API. That announcement also promised open weights for upcoming models. The October 7 release supplies two downloadable members of the family, rather than simply another way to access the hosted model.

The decision interface starts with an input, such as a customer message, and questions that define the permitted answers. The model processes them in a single forward pass—one run through the model—and does not generate answer tokens, the chunks of text a generative model produces.

• Yes or no: ask whether a condition holds, such as whether a customer wants a refund.
• A choice: supply named options, such as billing, technical support or fraud, for the model to select among.
• A score: provide an ordered scale, such as urgency levels, rather than ask for a free-form assessment.

Several questions can share the same input. Liquid AI’s release code asks whether a duplicate-charge complaint requests a refund, which team should handle it, and how urgent it is. Those are decisions about the message, not a drafted reply to the customer.

Different foundations, different input limits

d1-3B accepts text and images and is built from LFM2.5-VL-3B, Liquid AI’s vision-language model. The smaller d1-omni-600M uses a different foundation, LFM2.5-Encoder-350M, with added image and audio components. It accepts text with images or text with audio; the release does not describe all three arriving together.

Liquid AI labels the smaller model an early research release still under development. It publishes no speed results for that version. The device-level latency figures therefore apply to d1-3B, not to both new downloads.

16 msJetson AGX Thor

Liquid AI reports 16 milliseconds for one question on Jetson AGX Thor.

26 msJetson AGX Orin

Liquid AI reports 26 milliseconds for one question on Jetson AGX Orin 64 GB.

50 msJetson Orin Nano

Liquid AI reports 50 milliseconds for one question on Jetson Orin Nano.

The workload changes the timing substantially. On the Thor, three questions took 20 milliseconds, while a 3,400-token input took 220 milliseconds. A 384-pixel image took 35 milliseconds. The single-question headline is useful, but it is not a blanket response time for longer inputs or image tasks.

Text scores leave the multimodal test unfinished

Across seven public datasets, Liquid AI reports a mean score of 82.9 for d1-3B, compared with 81.1 for Decider 4B. The smaller d1-omni-600M scored 78.4, versus 77.1 for Decider 2B. The tests cover reading comprehension, toxicity detection, intent classification, medical questions and understanding across languages.

The company also reports a Decision Index 0.2.1 score of 48.57 for d1-3B, ahead of Decider 35B-A3B at 47.11. These are Liquid AI’s evaluations. Its release supplies no vision or audio benchmark scores, explaining that Decision Index v0.3 has only a private vision split and audio decision benchmarks remain an open problem.

For developers ready to evaluate the models, the release includes a Transformers loading example requiring version 5.14 or later and trust_remote_code=True, because the models ship their own code. Liquid AI also offers System One Arcade demos. Downloads are available now; broad published evidence for their image and audio decisions remains unfinished.