Ultramarine Flash v1: scaling nonverifiable RL (www.paperinstruments.com)

Ultramarine 1 Flash has been introduced as a new agentic model leveraging large-scale reinforcement learning (RL) on the open base model GLM-5.3 Flash. Demonstrating significant advancements, Ultramarine achieves an average improvement of 4.1 points across six knowledge-work benchmarks, while also generating 6.5% fewer tokens. It operates at a cost of $0.13 per task, outperforming competitor models like Claude Fable 5 and GPT-5.6 Sol, indicating its efficiency and efficacy in professional contexts that often involve complex, non-verifiable tasks such as drafting legal documents or financial reports. This development is crucial for the AI/ML community as it shifts focus from verifiable tasks, which have dominated prior research, to the non-verifiable realm where quality is often subjective and cannot be easily measured by conventional means. Ultramarine formalizes the RL task synthesis for these open-ended interactions, creating an effective curriculum that addresses the unique challenges of grading diverse outputs. By employing a multi-faceted curriculum that spans various professional domains and adopting synthetic data generation methods, Ultramarine aims to enhance the training and reliability of AI models in knowledge work, paving the way for more intuitive and capable agents in real-world applications.