Ultramarine 1 Flash: scaling RL for non-verifiable knowledge work, +4.1 pts
lmoroney · x · 2026-10-15
Paper Instruments released Ultramarine 1 Flash, an agentic model for professional knowledge work, post-trained with large-scale RL on GLM-5.3 Flash.
- Challenge: finance, law, consulting and operations tasks can't be graded by unit tests or regex; quality hinges on judgment, so rewards must be shaped per task via rubrics.
- Method: synthetic tasks with environments, instructions, and binary rubric criteria that name the substantive conditions earning credit, used to grade Excel, PowerPoint, and Word outputs; curriculum spans 4 domains, 19 practice areas, and 88 environment types.
- Results: +4.1 points average across six knowledge-work benchmarks with 6.5% fewer tokens; $0.13 per task, beating Claude Fable 5 ($8.19) and GPT-5.6 Sol ($2.09), nearing Claude Opus 5.5 and GPT-6 Sol.
- Availability: live on their API now, weights in 2 weeks. The author also suggests an exercise for RL teachers: write three binary rubric criteria for an open-ended assignment — the hardest part of non-verifiable RL.
Related event: Paper Instruments Releases Ultramarine 1 Flash(2 posts)→
More from Models
- StepFun releases Step-5-Preview: 600B/27B-active flagship with 1M context — ResearchCrafty1804 · 2026-10-15
- UltraBERT: open 150M/380M encoders with 32K context and an unusual dual-masking recipe — algo_diver · 2026-10-15
- DeepSeek V4.1 measures its own decode speed at 430-460 tok/s — then the claim falls apart — teortaxesTex · 2026-10-15
- OpenCode offers $60 of Gemini 4 Argon usage for $10 until Nov 30 — m4rkmc · 2026-10-15
- After losing 5 Claude accounts, one user's controlled tests show what actually triggers bans — jiayuan_jy · 2026-10-15
- Dev Says Jev Could Easily Detect X Tweet Spam, But X Hasn't Prioritized It — ivan_bezdomny · 2026-10-15