The LLM Data Company

Paper Instruments — formerly The LLM Data Company — is an AI research and tooling company betting on open-source frontier models fine-tuned for knowledge work. Instead of chasing general-purpose AGI, they train specialized models inside production harnesses, using autoresearch to curate tasks and rewards for on-policy reinforcement learning. This approach has already produced Kos-1 Lite, a state-of-the-art medical model, and Kos-1 Experimental, which scaled environment-free RL to 1T parameters on Kimi K2.5. Their product line includes Paper Office, an agent-first office document Python library that lets agents create and edit docs programmatically, and Feather, an agent harness for knowledge work that's coming soon. Inkwell, their frontier model series, is also in the works. Open-source models and research are available now, so teams can experiment with the methodology even before the full tooling ships. On the research side, they've released DiligenceBench, an agent-first benchmark for evaluating long-form equity research, and partnered with Perplexity on DRACO, a frontier eval for advanced deep research. They've also published notes on rubric judge training, a methodology for LLM evaluation that improves reliability. For enterprises in regulated domains like healthcare and finance, Paper Instruments offers a cost-effective alternative to generalist frontier models. Their domain-specific models, trained on production data, can cut serving costs while improving accuracy on narrow, high-stakes tasks. The website is minimal, but the open-source releases let you verify the claims hands-on. If you need immediate, off-the-shelf generalists, look elsewhere — but if you have the harness and the data, this is a specialty bet worth testing.