Liquid AI’s decision models split the choice between local control and hosted capability

Liquid AI’s d1-3B can run locally on small hardware, while its newer d1-ultra is offered through an API. An OrcaRouter comparison puts the choice in practical terms: pay for hosted capability, or take on the hardware and engineering work of running a smaller model yourself.

Watch Desk analysis

What happened

OrcaRouter says d1-3B has 3.12 billion parameters and that a 2.87GB quantised build fits on a Jetson Orin Nano. Its comparison lists a 50ms forward-pass time on that device. The article says d1-ultra is API-only and costs $0.042 per million input tokens; Liquid AI has not published a latency figure for it, according to OrcaRouter.

The models handle decision tasks, such as classifying an image or scoring a case, rather than writing free-form replies. OrcaRouter’s comparison says d1-3B’s weights are available under Liquid AI’s LFM Open License, while d1-ultra’s architecture and performance details are much less fully documented. The article also cautions that the models’ reported benchmark scores come from different evaluation setups, so they are not a clean head-to-head result.

Why it matters

The distinction is useful for anyone choosing where a decision model should run. Local inference can keep data on-device and avoid an API bill, but brings hardware, power and maintenance costs. A hosted model shifts that work to a provider and may suit harder cases, but its latency and practical performance need measuring in the intended pipeline.

The article’s example costs show why workload size matters: a modest number of hosted calls may cost less than buying hardware for the job, while very high volumes can make per-call pricing add up. The right answer depends on the task and scale, not simply which model has the more impressive score.

Our read

This is a worthwhile comparison because it turns “local versus cloud” from a slogan into a deployment decision. Start with d1-3B if data locality, offline use or call volume is the constraint; consider d1-ultra when the decision itself is harder. But test latency and accuracy on your own cases before trusting either with a consequential gate. A spreadsheet is a useful colleague, not an oracle.

What to watch

• Whether Liquid AI publishes d1-ultra latency and fuller technical details.
• How both models perform on the same decision tasks and evaluation setup.
• Whether local hardware and operating costs change the break-even point for real workloads.

Discussion spark: For a decision model handling sensitive or high-volume work, would you favour local control even if it means maintaining the hardware, or pay for a hosted model with less published performance detail?

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.