Open-source frontier-class model for knowledge work, post-trained from GLM-5.3 Flash by a five-person team for about $300k. It matches or approaches much larger closed models on internal knowledge-work benchmarks at roughly 60x lower serving cost, with API access now and weights coming soon.
Ultramarine 1 Flash is an agentic model for professional knowledge work, adapted from GLM-5.3 Flash using reinforcement learning. Paper Instruments describes training tasks across finance, law, consulting, and operations, and evaluates the model on knowledge-work and coding benchmarks. It can be accessed through Paper Instruments’ API using OpenAI-compatible Chat Completions and Responses formats, or the Anthropic Messages format. The model is also available as downloadable weights on Hugging Face for local serving; the model card documents a SGLang setup and recommends running it in an agent harness.
SGLang is documented as a serving engine for running Ultramarine 1 Flash locally.
