Paper Instruments released Ultramarine 1 Flash. It applies large-scale reinforcement learning to GLM-5.3 Flash. It improves by 4.1 points on average across six knowledge-work benchmarks. It generates 6.5% fewer tokens. It costs $0.13 per task on internal tests. It outscores Claude Fable 5 and GPT-5.6 Sol. It nears Claude Opus 5.5 and GPT-6 Sol. It targets non-verifiable professional tasks. It spans finance, law, consulting, and operations. It uses synthetic data to avoid production trace limitations. It relies on custom rubrics for grading open-ended work. It prevents mode collapse via combinatorial task diversity. It includes 88 environment types across 19 practice areas.
What changed
Verifiable tasks dominated agentic RL training, with coding as the largest share of the mix.
What it unlocks
Scaling non-verifiable reinforcement learning for open-ended professional knowledge work across finance, law, consulting, and operations.
- 4.1 points average improvement
- 6.5% fewer tokens
- $0.13 per task cost
- 88 environment types
What you need to act on it
- GLM-5.3 Flash base model
- synthetic task and environment generation pipelines
- custom rubric-based grading system
Sources