Subquadratic launches SubQ, a sub-quadratic LLM supporting 12M-token context at 1/5 the cost of other leading models

SubQ is built on a fully sub-quadratic sparse-attention architecture designed for 12M-token reasoning. The model processes only relevant relationships between words rather than all possible relationships, reducing attention compute almost 1,000× at 12M tokens.

SubQ operates at 150 tokens per second and costs 1/5 of other leading LLMs. It is positioned as a leader in long-context retrieval and coding tasks, with an API supporting 12M token context windows, streaming, tool use, and OpenAI-compatible endpoints.

A coding-agent plugin is available that claims ~25% lower cost and 10× faster exploration, with auto-redirect of expensive model turns. Early access is available via private preview.