Subquadratic Review on Quasa: The First Sub-Quadratic LLM

#Quasa #QUA #subquadratic

Subquadratic is a frontier AI research and infrastructure company building a new class of large language models. Their flagship model, SubQ, is the first LLM built on a fully sub-quadratic sparse-attention architecture. Unlike traditional Transformers that scale quadratically with context length ($O(n^2)$), SubQ scales linearly ($O(n)$), making multi-million token reasoning practical and efficient.

SubQ is designed specifically for long-context tasks. It can process entire code repositories, long agent histories, months of pull requests, and persistent states in a single prompt without quality degradation. The model currently supports context windows up to 12 million tokens and delivers strong performance on both long-context retrieval and general reasoning benchmarks. At 1 million tokens, SubQ requires dramatically less compute (claimed 64.5× less than dense attention) and runs significantly faster than standard attention mechanisms.

The company focuses on the principle that “efficiency is intelligence.” By solving the quadratic bottleneck at the architecture level, Subquadratic aims to enable a new generation of AI agents that can truly work with large amounts of context instead of relying on chunking, summarization, or retrieval hacks.

Subquadratic is ideal for developers, coding agents, enterprise teams, and researchers who need reliable long-context reasoning — especially for full-codebase analysis, long-running agents, and complex multi-document tasks.

Highlights

• First fully sub-quadratic (linear-scaling) LLM architecture.
• Multi-million token context (up to 12M tokens).
• Strong long-context retrieval with high efficiency.
• Designed for full repositories and persistent agent state.
• OpenAI-compatible API and coding agent integrations.

• Potential Considerations

• As a frontier research model, broader ecosystem and size variants are still expanding.
• Best suited for long-context use cases rather than general short-prompt chat.

Overall Verdict: 4.7/5 stars

Subquadratic is tackling one of the most important limitations of current LLMs — the quadratic cost of attention. By delivering practical multi-million token reasoning at far lower compute, SubQ represents a meaningful architectural step forward for agents and long-context applications.

Earn QUA reward by reviewing on Quasa.io too!

Get started: https://quasa.io/projects/subquadratic