Ultramarine 1 Flash

About Ultramarine 1 Flash

Ultramarine 1 Flash is an agentic AI language model developed by Paper Instruments for professional knowledge work. Announced on October 14, 2026, the model is built on GLM-5.3 Flash and uses reinforcement learning to improve its ability to complete complex, open-ended professional tasks. Its training focuses on work requiring analytical reasoning, document interpretation, judgment, and the preparation of structured business deliverables.

The model's reinforcement learning curriculum covers four professional domains: finance, law, consulting, and operations. Paper Instruments developed synthetic tasks across 19 professional practice areas and 88 environment types. These environments simulate workplace scenarios involving financial models, contracts, regulatory filings, business correspondence, operational records, and connected enterprise systems. The training approach uses task-specific evaluation rubrics to assess professional outputs that cannot be verified through conventional automated tests alone.

Ultramarine 1 Flash was trained using Feather, a Python-based agent harness developed by Paper Instruments. Feather provides tools for reading and modifying files, searching workspace content, running shell commands, managing background processes, and tracking task progress. The training environment also uses Paper Office libraries for working with Word documents, Excel spreadsheets, and PowerPoint presentations. These capabilities are part of the documented agent environment rather than standalone features guaranteed by the model without appropriate tools.

The published training configuration specifies a 131,072-token context window and a maximum of 250 agent turns. Paper Instruments reports results across professional knowledge-work benchmarks and coding evaluations, including JobBench, FrontierFinance, AutomationBench, DiligenceBench, APEX-Agents, Terminal-Bench, and DeepSWE. The model is intended for developers and organizations building agentic applications that involve research, business analysis, document workflows, and multistep professional tasks.

Key Features

• Agentic Knowledge Work: Trained for complex professional tasks involving finance, law, consulting, and business operations, including document preparation and analysis.

• Reinforcement Learning: Uses a synthetic reinforcement learning curriculum with task-specific grading rubrics designed to evaluate open-ended professional deliverables.

• 131,072-Token Context Window: The published training configuration specifies a 131,072-token context window for processing substantial task instructions and workspace information.

• Tool-Based Workflows: Trained using the Feather agent harness, which provides file operations, workspace search, shell execution, process management, and task planning tools.

• Document Processing Workflows: Training incorporates Paper Office libraries for editing and working with Word, Excel, and PowerPoint files in agent environments.

• Reasoning and Coding Evaluations: Evaluated on six professional knowledge-work benchmarks and two coding benchmarks, with published results for reasoning-intensive and software engineering tasks.