Open-source 149M LateOn decision model ditches classifier heads for late interaction
antoine_chaffin · x · 2026-09-29
tasksource-jev-nano-v0 is now open on Hugging Face: a 149M-parameter decision model built on the LateOn architecture, Apache-2.0 licensed.
- What it does: picks the best option from a candidate list (fraud routing, intents, topics, sentiment) using token-level late interaction instead of a classifier head
- Training: fine-tuned on 512k real decisions across three judgment types (choice, yes/no, graded score), no teacher model
- Architecture perks: the situation is encoded once and cached across decisions; outputs are exactly permutation-equivariant regardless of option order
- Scores: 46.8–49.7% on unseen Tasksource tasks, 44.7% on Banking77, 57.5 micro on classifier-benchmark v2
A notable lightweight option for cheap multi-option decision tasks in constrained settings.
More from Models
- OpenAI vs Anthropic model race: 6.1 Astra "tomorrow or bust" as Fable 5.5 looms — karmay007 · 2026-09-29
- OpenAI may unveil a "guardian" feature at Dev Day, speculates AI insider — imjustnewatai · 2026-09-29
- Dev hands-on: new Claude model is chattier than Opus but far less than Sonnet 5 — felixrieseberg · 2026-09-29
- NaiveAILab Confirms Its New Model Is Built on MiMo-V2.5, Sparking Open-Source Debate — teortaxesTex · 2026-09-29
- Sentdex has run 4B+ tokens locally on GLM 5.3 Flash — his most-used local model ever — Sentdex · 2026-09-29
- Leaked benchmark shows Sonnet-5.5 above Astra, source unverified — adonis_singh · 2026-09-29