Paper Instruments trains Ultramarine 1 Flash with reinforcement learning on work that has no single right answer

Paper Instruments has introduced Ultramarine 1 Flash, an agent model for professional knowledge work in areas such as finance, law, consulting and operations. Its technical report describes large-scale reinforcement learning on work that cannot be checked automatically. Coding answers can be tested with unit tests or simple pattern checks, but tasks like diligence memos or litigation chronologies cannot, so the reward has to be designed for each task. The company started from the open model GLM-5.3 Flash. It says Ultramarine improves on that base across all six knowledge-work benchmarks it tested, by 4.1 points on average, while producing 6.5 percent fewer tokens. On the company's internal suite, Ultramarine costs $0.13 per task. The company says it outscores Claude Fable 5 ($8.19) and GPT-5.6 Sol ($2.09) and comes within three points of Claude Opus 5.5 ($6.96) and GPT-6 Sol ($1.26), with all four rivals set to maximum reasoning. The model does not lead everywhere. On the company's own charts, Gemini 3.8 Flash and Claude Fable 5 score higher on DiligenceBench and APEX-Agents. All of the results are the company's own. Ultramarine is available through an API, and the company says model weights are "coming soon."