Liquid AI open d1: when a fast decision beats an AI answer

Liquid AI released two open-weight d1 decision models on October 7, 2026: d1-3B and the smaller experimental d1-omni-600M. These are not general-purpose assistants built to write long explanations. Their core purpose is to read a state, evaluate a fixed question or set of candidate outcomes, and return probabilities without generating an answer token by token. That design is relevant to teams that repeatedly classify support tickets, route leads, inspect images, or make on-device decisions with strict latency limits.

What makes a decision model different?

A typical language model may respond to a classification task with a paragraph or a JSON object that must be parsed and checked. Liquid's d1 architecture instead handles the decision in a single forward pass, returning probabilities across the allowed outcomes. For a workflow where the only valid answers are, for example, 'sales', 'support', 'spam' or 'human review', the model does not need to spend time composing a written explanation. The trade-off is that the outcome set and the business logic need to be specified well in advance.

The two releases and their limits

Both checkpoints are available as open weights through Hugging Face, with local inference support including llama.cpp. Liquid's release describes support across Apple, AMD, Qualcomm and NVIDIA environments. These are model distribution and implementation claims, not a guarantee that every device configuration achieves the same throughput or that deployment has no licensing, integration or maintenance obligations.

What the published benchmarks actually say

Liquid reports a Decision Index v0.2.1 public-split score of 48.57 for d1-3B and 15.95 for the d1-omni-600M experimental checkpoint. Across seven text benchmarks presented in the announcement, their reported averages are 82.9 and 78.4 respectively. The company also reports an 8 ms single-question response for d1-3B on an RTX 4090, 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin and 50 ms on an Orin Nano. These are vendor-reported measurements on specified setups, not independently reproduced end-to-end application benchmarks.

A real workflow: lead and support routing

1. Define the closed set of outcomes: qualified sales inquiry, customer support, spam and human review.
2. Build a labeled evaluation set from actual inquiries, removing private identifiers and sensitive fields where feasible.
3. Run the decision model and compare its predictions with a simple rules-based baseline and an existing generative model.
4. Measure false positives, false negatives, calibration and time per decision; assess whether different errors have different business costs.
5. Choose a conservative confidence threshold. Send uncertain or high-impact cases to a person instead of forcing an automatic decision.
6. Only automate routing after monitoring, logging, versioning and rollback are established. Reevaluate the system when inquiry patterns change.

Where the open-weight release matters

Open weights make it possible to examine, adapt and deploy a checkpoint closer to the data source, including on devices where sending every input to a hosted model is undesirable. For robots, cameras, field devices and local business systems, network reliability and responsiveness can be as important as benchmark scores. The smaller omni model is particularly interesting as an experimental route to combined text-and-image or text-and-audio decisions, but Liquid explicitly labels it an early checkpoint. The release says dedicated audio decision benchmarks are not yet mature.

• Test the exact input format: text, text plus image, or text plus audio rather than assuming one model supports all combinations equally.
• Confirm hardware latency with representative payloads, not only a short single-question benchmark.
• Review deployment and model license terms before commercial rollout, and document what version is deployed.
• Keep a rule-based or human fallback when the model is uncertain, unavailable or sees unfamiliar data.
• Protect sensitive input and audit decisions that affect people, access or money.