News · 2026-10-08

Liquid AI releases d1 models that choose answers without generating prose

Liquid AI released d1-3B and d1-omni-600M on October 7 as open-weight models for decisions rather than conversation. They take a state and named questions, then return typed answers in one forward pass instead of generating prose token by token. That makes them potential components for routing, moderation and inspection inside larger agents, with task accuracy and license conditions still requiring scrutiny.

Key facts

• Liquid reports 8-millisecond single-question latency for d1-3B on an RTX 4090.
• The October 7 release contains a text-and-image model and an experimental smaller multimodal model.
• Commercial use at annual revenue of $10 million or more needs a separate license under the license text.
• The primary source is Liquid AI’s Open d1 announcement.

Liquid titles the release “Open d1.” The important distinction is functional. A support application might supply a message and ask three questions: does it request a refund, which department should receive it and how urgent is it? The model can answer the questions over the application’s chosen options in one pass. A chatbot would generally have to construct a string explaining or formatting those answers before the application extracted them.

A concrete analogy is a sorting desk beside a writer. The sorter applies labels and directs work to the right tray. The writer handles the longer explanation or final judgment. Neither role is inherently more valuable; the system benefits when it avoids paying for a paragraph where a bounded choice is sufficient. The existing discriminative-model lesson explains that distinction in more detail.

The two checkpoints have different designs. The d1-3B model card describes a model accepting text and images. Liquid says its training combined related backbone weights, varied training seeds and data mixtures, and used answer-option shuffling and shortcut-resistant data. These details matter because a choice model can learn superficial associations with option order or wording rather than the decision itself.

The d1-omni-600M card describes 587 million parameters spread across a shared trunk and decision head, a vision encoder and a speech encoder. It accepts text with an image or text with up to 30 seconds of speech. Those are alternative modes in the documented interface. Its audio training targets English speaker–assistant requests, including topic and intent. A small model name should not become a claim of universal multilingual or simultaneous audio-video understanding.

Performance evidence is strongest where Liquid names the device and task. Its d1-3B single-question measurements include 30 milliseconds on a Mac M5 Pro, 50 milliseconds on a Jetson Orin Nano and 8 milliseconds on an RTX 4090. The latter is a 24 GB graphics card; that is the measured hardware, not a stated minimum memory requirement. The reviewed sources do not specify a universal runtime video-memory requirement, and d1-omni has no published latency measurement in the release. A verified checkpoint download-size figure is also absent from the dossier, so none is inferred from parameter counts.

Liquid reports strong results for d1-3B on its Decision Index and mixed task-level results across public text benchmarks. These are company-run comparisons. They do not prove that a small decision model can replace every larger model, or that its probability for an option is reliably calibrated in a new application. A deployment should test its real labels, ambiguous cases and out-of-distribution inputs. The lesson on calibration explains why confident scores and dependable decisions are separate questions.

The local path is tangible. Liquid documents Transformers loading, and says d1-3B has llama.cpp support across several hardware families. The official card’s use of remote model code also makes code provenance part of installation. Its linked hosted camera demos provide a convenient trial, but a hosted page is not proof that inference runs on the visitor’s device. The supplied feed mentions a community WebGPU demonstration; no official Liquid browser build or phone measurement was established.

The licensing boundary deserves equal prominence. Both model cards point to the LFM Open License. It is based on Apache terms but adds a commercial condition at annual revenue of $10 million or more. Liquid’s FAQ uses slightly different boundary wording, so the actual license text is the safer basis for deployment decisions. Qualified nonprofit research conditions are not a blanket exemption for commercial research. Open weights describes availability of the model files, not unrestricted permission for every use.

The release gives developers a credible reason to separate small repeated decisions from large generative calls. Model routing is one obvious use, but visual checks and speech intent are other bounded applications. The honest caveat is that d1-omni is experimental, measurements remain vendor-reported and browser or phone claims outrun the official evidence. The value will come from a specific reliable decision on a specific device, rather than the model’s small size alone.

Primary source, verified: read the paper →

Ground Truth. (2026, October 8). Liquid AI releases d1 models that choose answers without generating prose. Ground Truth. https://groundtruth.day/news/liquid-d1-one-pass-decision-models.html

@misc{groundtruth:liquid-d1-one-pass-decision-models,
  title  = {Liquid AI releases d1 models that choose answers without generating prose},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/liquid-d1-one-pass-decision-models.html}
}

Topics: open-weights · liquid-ai · classification · edge-ai · agents