Paper Instruments has released Ultramarine 1 Flash, an open model trained on professional knowledge work using a novel reinforcement learning approach. This method, called task-rubric co-design, uses synthetic data and detailed rubrics to grade non-verifiable tasks, overcoming limitations of traditional RL and expensive human feedback. The model achieves Fable-class performance on knowledge work benchmarks at a significantly lower cost than existing models like Claude Fable 5 and Claude Opus 5.5, potentially making AI assistance more accessible for fields such as law, finance, and consulting, and enabling AI companions capable of nuanced emotional intelligence and creative output. AI
IMPACT This new training method could make AI assistance more affordable for professional services and enable more nuanced AI companions.
RANK_REASON The item describes a new open model release and a novel training methodology for AI, positioning it as a research advancement. [lever_c_demoted from research: ic=1 ai=1.0]
- Claude Fable 5
- Claude Opus 5.5
- GLM-5.3 Flash
- OpenClaw
- reinforcement learning from human feedback
- Ultramarine 1 Flash
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
