the short answer
Julia 1 is a 144.3M-parameter model from Supersonic Labs, with Apache 2.0 weights, Python inference and a separate ONNX/WebGPU artifact. It may suit a CPU or browser deployment where Jev’s hosted route does not. The publisher’s benchmark card uses supplied Jev reference figures rather than a fresh paired run. Test Julia’s language coverage, 2–20 option limit and calibration against Jev on your own labeled cases.
- Publisher
- Supersonic Labs; independent of TypeSafe AI
- Architecture
- 144.3M-parameter model built from mmBERT-small
- Artifacts
- Apache 2.0 weights and Python runtime; separate ONNX/WebGPU export
- Native option limit
- 2–20 options per Choice or Score request in the published runtime
- Benchmark status
- Publisher-run; Jev figures are supplied references, not a new matched run
What Fits on a Local Machine
Julia 1’s model card describes a multilingual mmBERT-small encoder that scores answers you supply. The Python predict(state, questions) path accepts Choice, Score and Noul-shaped questions. A separate ONNX artifact includes a WebGPU adapter. Both artifacts come from Supersonic Labs.
The runtime accepts up to 8,192 combined tokens and 2–20 options per question. Its published accuracy tests used 1,024 tokens, so the longer accepted input has no matching quality claim. A hierarchical router can split larger answer lists, though that produces a different distribution from one question over every original option.
Test a Local Routing Need
Take a ticket where sign-in blocks an invoice download. Account access is the expected first owner under this rubric. Whether Julia or Jev finds it is the test. Add ambiguous tickets and each language your queue receives, then report the errors separately.
| Question | Matched fixture |
|---|---|
| State | A customer ticket with the account facts allowed for routing |
| Choice | First blocker: account access, billing, technical or review |
| Environment | Record whether Julia runs on CPU or WebGPU; record Jev provider and region |
| Outcome | Compare against blinded human routing labels and later resolution |
| Failure | Count overflow, missing data, timeout and unsupported hardware separately |
What the Julia Benchmarks Can Establish
Supersonic Labs reports AG News, Emotion, Banking77, MASSIVE locale and typed-decision results. The classification pilots have 100 examples each, and the typed suite has 400 cases. The Jev figures were supplied from another protocol. Those columns do not show what would happen if both models saw the same new tickets.
The card calls out knowledge, multi-step reasoning, ambiguity and long candidate lists as weak spots. Put each in the test set. If local inference is the draw, count CPU or browser memory, cold starts and maintenance alongside the hosted Jev bill.
How to Benchmark Jev and Julia 1
The paired benchmark method specifies what to save with each run. For deployment planning, the self-hosting guide covers Julia-like local options and Jev’s hosted status.
- Pin Julia weights, Python or ONNX runtime, encoding settings and the Jev model and route.
- Use the same labeled tickets and equivalent answer definitions within Julia’s option limit.
- Check native response shape and all failures before normalizing results.
- Measure class errors, language slices, calibration, review coverage and end-to-end latency.
- Report hardware and full operational cost; set action thresholds separately.
FAQ
Is Julia 1 made by TypeSafe AI?
No. Supersonic Labs publishes Julia 1 as an independent model and runtime.
Can Julia 1 run without a GPU?
Its model card documents CPU inference and a separate ONNX/WebGPU artifact. Verify the runtime on your target hardware.
Is Julia 1 more accurate than Jev?
The card does not report a new matched Jev run. Compare both on the same held-out labels before drawing a task-specific conclusion.
Sources
Checked against the sources below on October 2, 2026. Model versions, prices and limits change.
- Supersonic Labs: Julia 1 model card
- Supersonic Labs: Julia 1 ONNX and WebGPU
- TypeSafe AI: Jev primitives
- TypeSafe AI: Jev models
Change note: Added Julia 1 provenance, serving options and benchmark evidence boundaries.