Ultramarine 1 Flash report: rubric-based task synthesis for non-verifiable RL
lmoroney · x · 2026-10-15
Paper Instruments published the technical report for Ultramarine 1 Flash (same story as the earlier thread; this is the source article).
- Post-trained with large-scale RL on GLM-5.3 Flash, targeting non-verifiable professional knowledge work in finance, law, consulting, and operations.
- Method: formalize each task like a verifiable one — environment, instruction, and verification system — with per-domain synthetic pipelines and custom rubrics grading Excel/PowerPoint/Word artifacts.
- Costs $0.13 per task, averages +4.1 points across six benchmarks with 6.5% fewer tokens; live on their API, weights in 2 weeks.
Related event: Paper Instruments Releases Ultramarine 1 Flash(2 posts)→
More from Models
- Claude Code's usage-limit screen is how you scare a developer — Arindam_1729 · 2026-10-15
- Liquid AI releases LFM2.5-24B: just 2B active params per token for local agents — helloiamleonie · 2026-10-15
- "People wildly underestimate how much model writing style matters," argues viral AI post — Promptmethus · 2026-10-15
- Ultramarine 1 Flash launch goes viral as every hype post carries a "Paid partnership" tag — suchenzang · 2026-10-15
- Ling-3.1-flash hits Hugging Face with thinking toggle and tool-call template — AdinaYakup · 2026-10-15
- Apertus 1.5 extends open models to multimodal reasoning with 262K context — swiss-ai · 2026-10-15