Supersonic Labs Releases CPU-Friendly AI Router With a Weakness on Crowded Choices

A router that runs on a CPU still has to pick the right destination. Supersonic Labs released Julia 1 on September 26 with downloadable weights and code, but its own banking test found a sharp limit: the model got 64 of 100 examples right when choosing among similar categories. The company cited an 87% reference score for that test. Julia 1 offers a way to make text decisions locally; its release also gives developers a reason to test those decisions before relying on them.

An answer from a supplied menu

Julia 1 is a 144.3-million-parameter model built on mmBERT-small, an existing multilingual model for processing text. It takes context, a question and two to 20 possible answers, then scores the supplied choices. It can classify a message, send it to a category, assign an ordered score or answer yes or no. It does not write a free-form response.

That narrower job shapes the release. A developer supplies the options rather than asking the model to invent an answer. Supersonic’s example sends a customer message about a duplicate charge to a category such as Billing; another question could assign its urgency. Those are separate decisions using the same input format, not a claim that the model can resolve the customer’s problem.

Developers can download the weights and Python interface from Hugging Face under Apache 2.0. Supersonic says the evaluated setup accepts up to 1,024 tokens across the context, question and options. A router can narrow larger sets of candidates in groups, but that step creates a risk: it may discard the right answer before the final choice.

The tests split by task

The banking result was not typical of every test Supersonic published. Julia 1 got 94 of 100 items right in a four-label news pilot and 86 of 100 in a six-label emotion pilot. It also answered 1,463 of 2,000 typed decisions correctly, or 73.15%, just above the supplied 72.70% Jev reference. These are company-reported evaluations, not a promise of the same accuracy on a developer’s messages.

The banking pilot asked the model to distinguish among 72 labels, after first narrowing the list. That differs substantially from choosing among four news labels. Supersonic identifies long lists of similar categories as an area for improvement and says the narrowing step can lose the correct answer. A team using Julia 1 to route requests would need to test its own labels, especially when their meanings overlap.

A broader multilingual test adds another measure, with a different task. On MASSIVE, Julia 1 chose among 18 scenarios and got 110,573 of 154,648 cases right across 52 locales, or 71.50%, according to Supersonic. The result shows the scope of that evaluation; it does not settle how well the model will distinguish a particular organization’s categories.

CPU-ready does not mean equally quick on every job

Julia 1’s weights occupy 550.5 MiB. Supersonic measured it on an Apple M4, an Intel Core i5-1235U computer and a Samsung tablet. The tablet completed 40 decisions in eight seconds, or five per second, in a run that fell back to CPU execution. That demonstrates a possible local deployment, not a speed a developer should expect on every device.

The Intel results show why the workload matters as much as the processor. Supersonic recorded a median of 294.81 milliseconds for typed decisions but 3,713.54 milliseconds for the Banking77 pilot, which includes narrowing 72 categories before selection. Those are different tasks, not two timings for the same request. Anyone putting the model in a live routing path would need to measure the full path with their own inputs.

In a separate CPU evaluation, Julia 1 got 1,451 of 2,000 typed decisions right, or 72.55%. On Banking77, it got 60 of 100 right and abstained three times; abstentions counted among the 100 cases. The earlier 64% banking result and this 60% CPU result come from separate reported runs. Neither should be treated as a fixed accuracy rate for a new application.

Available weights, planned service

The downloadable model is the current release. Supersonic says it is still building an API and plans to charge $0.025 per million input tokens, with no charge for output tokens, when access opens. That price is a plan, not a service developers can use today. The company also says Julia 2 remains in development and is intended to use its own foundation architecture rather than mmBERT-small.

For now, Julia 1 gives developers an open model to run and inspect. Its clearest test is not whether it starts on a CPU, but whether it picks reliably among the choices an application actually gives it. Supersonic’s banking results make that question especially pressing for routes with many similar destinations.