Julia-1: a 144M-param classifier that runs on CPU, wins 3 of 4 tasks
juliusadml · x · 2026-09-27
Supersonic Labs released Julia 1, its first compact decision model built on mmBERT-small. At 144.3M parameters it runs on a CPU, taking a question plus 2-20 candidate answers and picking one with scores. In a Sept 24, 2026 eval it beat the Jev reference on 3 of 4 tasks: 1,463/2,000 typed decisions (73.15%, +0.45pp), 94/100 news classification, and 86/100 emotion classification (+38pp). A 72-category banking pilot was its failure: 64/100 vs the reference's 87%, with long lists of similar categories the weak spot.
More from Models
- banteg: astra finds nothing to flag after Opus 5.5, unlike nitpicky prior models — banteg · 2026-09-27
- Opus 5.5 Impresses Early Users With High Capability per Token and Few Usage Limits — signulll · 2026-09-27
- Burkov mocks Google AI Mode: it answers from memory instead of actually searching — burkov · 2026-09-27
- OpenAI Codex agents went rogue, spawned 826 parallel tasks and burned ~$78,000 — lorenzomassaro · 2026-09-27
- 17M-parameter model beats frontier LLMs 80% of the time after cheap synthetic-data finetuning — max_paperclips · 2026-09-27
- goodside demos Claude Opus 5.5 (Extra) quickly explaining Primer (2004) — goodside · 2026-09-27