Liquid AI has released two open d1 decision models, including one it says can answer a question in under 50 milliseconds on each of four tested NVIDIA devices. The models return structured decisions in a single pass rather than generating text, offering developers a smaller, faster option for classification and routing tasks.
Watch Desk analysis
What happened
Liquid AI says d1-3B accepts text and images, while the experimental d1-omni-600M accepts text with either images or audio. Both are open-weight and available on Hugging Face. The company’s release notes include model details and example code; using the models requires transformers>=5.14 and loading their supplied code with trustremotecode=True.
On the Decision Index 0.2.1, Liquid AI reports that d1-3B scored 48.57, ahead of the other listed models under 10 billion parameters. Across seven public datasets covering tasks including intent classification, toxicity detection and medical question answering, the company reports mean scores of 82.9 for d1-3B and 78.4 for d1-omni-600M. These are the release’s benchmark results, not an independent comparison.
Key findings
- Fast responses on edge devices
Liquid AI reports d1-3B answered one question in 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin and 50 ms on Jetson Orin Nano. - Text-and-image decisions
d1-3B is built on a vision-language model and accepts text and images; d1-omni-600M adds an audio input route. - One pass, structured outputs
The models return choices or scores rather than generated tokens, which suits tasks such as assigning a support ticket to a team. - The smaller model is still experimental
Liquid AI says d1-omni-600M is under further development and publishes no speed figures for it.
Why it matters
A developer handling lots of routine decisions may not need a conversational model to write an answer. A compact model that returns a category, choice or urgency score could be quicker to run locally, and the reported device timings make the edge-computing case concrete. Liquid AI also says three questions took only 1.3 times as long as one on AGX Thor, suggesting that batching decisions could be useful.
The evidence has boundaries. The release says the benchmark covers seven public datasets, but does not report vision or audio benchmark results. Its stated top ranking applies to the Decision Index 0.2.1 results, not a universal measure of decision quality. In other words, the numbers are a promising starting point, not a permission slip to route every tricky judgement through a tiny model.
Our read
This is a useful development for teams that need quick, structured decisions, especially where sending every request to a larger generative model would be unnecessary. Start with a narrow, measurable task and test the model against your own cases, including the awkward ones. The speed claim is appealing; whether the decisions are good enough for a real workflow is the test that matters.
What to watch
- Whether independent evaluations reproduce the reported Decision Index results.
- How d1-omni-600M performs as it moves beyond its experimental release.
- Whether Liquid AI publishes vision and audio decision benchmarks.
Discussion spark: For routine workflow decisions, would you trust a fast, single-pass model to act directly, or require a larger model or human review before anything consequential happens?
Sources and evidence
- Multimodal open d1 decision models for the edge (7 October 2026, 16:54 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.