A cutting-edge small AI Lab overseas, Paper Instruments, has a very small team and mainly focuses on post-training for large models. - Their niche is in vertical domains—knowledge work (like creating professional reports, financial models, legal documents, consulting deliverables, etc.) - Business model: open-source + inference API services - Technical focus: emphasizes agentic RL (reinforcement learning) in non-verifiable domains - Funding is relatively small at $4 million, categorized as an early-stage YC project (Y Combinator 2025 Spring (S25) batch company) On October 14, 2026, they released Ultramarine 1 Flash (post-trained on GLM-5.3 Flash). They claim it performs close to Claude Fable 5 on knowledge work tasks, at approximately 60x lower cost; total post-training cost was around $300,000. These post-trained small models not only exist within large companies like Coinbase but also in external service markets. This means some post-training teams can create models with performance and cost advantages over top-tier labs, offering open-source and API-based business models, carving out their own survival space. Their CEO x: Daanish Khazi @bertgodel More about their research on RL in non-verifiable domains: https://www.(paperinstruments.com)/blog/ultramarine