Subquadratic

About Subquadratic

Subquadratic is a frontier AI research company building a new class of models that scale efficiently. The company's main product is SubQ Snap, a long-context decision model designed for structured decisions in agentic workflows. It features a 384k-token context window and processes 100K tokens in less than a second at $0.04 per 1M tokens. SubQ Snap is built using Subquadratic Sparse Attention (SSA), a proprietary algorithm that isolates the tokens and relationships that matter, ensuring compute is used efficiently rather than processing every possible relationship between tokens. The company also offers SubQ 1.1 Small, a long-context language model currently in early access that can reason across millions of tokens in a single prompt.

• long-context
• ai
• agentic workflows
• language model
• efficient-ai
• decision model
• enterprise

FAQ

What is a decision model?

Decision models are AI models that evaluate supplied context against defined questions and return structured answers. They excel when deep reasoning or nuanced outputs of a frontier LLM are not required. They are best at routing, ranking, classifying, and yes/no calls, and output a choice, score, or label with structured, defined answers that require no parsing.

What is Subquadratic Sparse Attention?

Subquadratic Sparse Attention (SSA) is a proprietary algorithm that isolates the tokens and relationships that matter for computation, rather than processing every possible relationship between tokens like traditional transformer-based LLMs. This ensures compute is used efficiently.

How much does SubQ Snap cost?

SubQ Snap costs $0.04 per 1 million input tokens, with new users receiving 100M free tokens. There are separate costs for output tokens which are not specified on the homepage.

How fast is SubQ Snap?

SubQ Snap processes 100K tokens in less than a second, a speed advantage over frontier LLMs which would take minutes for the same task.

What is SubQ 1.1 Small?

SubQ 1.1 Small is a long-context language model currently in early access that can reason across millions of tokens in one prompt, handling entire repositories, whole artifacts, and long-running agent state.