Subquadratic, or SubQ, is a frontier AI research and infrastructure company that specializes in building a novel class of large language models (LLMs). Unlike many major laboratories that concentrate on iterative enhancements to existing Transformer architectures, Subquadratic is dedicated to pioneering fundamental changes at the core model architecture level. This innovative approach enables the development of large-context, multi-modal inference capabilities that scale with superior efficiency, surpassing the limitations typically encountered by traditional Transformers. SubQ is recognized as the first fully sub-quadratic LLM, specifically engineered for advanced 12-million-token reasoning. This capability allows AI agents to effectively operate across extensive and complex data environments, including full code repositories, comprehensive historical data, and persistent operational states, all without experiencing any degradation in quality or incurring prohibitively high computational costs. The company's foundational work is backed by researchers who previously contributed to leading institutions such as Meta, Google, Oxford, Cambridge, and BYU.

SubQ operates on a groundbreaking, fully sub-quadratic sparse-attention architecture, which marks a significant departure from the conventional Transformer models that typically expend computational resources by processing every conceivable relationship between words. This traditional method often leads to substantial inefficiency as only a small fraction of these relationships are truly material. In contrast, SubQ intelligently identifies and selectively focuses on only the most pertinent relationships, thereby drastically minimizing computational waste. This selective processing enables SubQ to manage extremely large contexts, such as those involving up to 12 million tokens, with remarkable efficiency and cost-effectiveness. The model is capable of processing data at an impressive rate of approximately 150 tokens per second, achieving this at roughly one-fifth the operational cost when compared to other leading large language models currently available in the market. This architectural innovation redefines how LLMs can scale to handle extensive data loads.

Some of the key features are:

• 12M Token Context Window: Provides an unprecedented capacity for reasoning across vast amounts of information in a single prompt, allowing for comprehensive data analysis.
• Sub-Quadratic Architecture: Leverages a novel sparse-attention design that processes only relevant word relationships, leading to an almost 1,000x reduction in attention compute at 12M tokens, significantly enhancing efficiency.
• High Accuracy: Demonstrates robust performance in demanding long-context retrieval and coding tasks, consistently achieving competitive or superior benchmarks when compared against other frontier models in the industry.
• Cost Efficiency: Operates at an exceptionally low cost, approximately one-fifth the expense of other leading LLMs for tasks requiring extensive context, making advanced AI more accessible.
• High Speed Processing: Capable of processing information at a rate of 150 tokens per second, facilitating rapid data analysis and response generation.
• Developer API: Offers a comprehensive, full-context API equipped with streaming capabilities, robust tool use, and endpoints that are fully compatible with OpenAI standards for seamless integration into existing development workflows.
• SubQ Code for Agents: Provides a specialized long-context layer meticulously designed for integration with coding agents, resulting in approximately 25% lower billing costs and a 10x faster exploration of codebases.
• Third-Party Validated Results: All benchmark performance results are independently verified by third parties, ensuring the reliability and trustworthiness of SubQ's stated capabilities.

SubQ achieves its unparalleled efficiency through its proprietary and innovative fully sub-quadratic sparse-attention architecture, which radically deviates from the quadratic complexity (O(n²)) inherent in standard Transformer models. This architectural breakthrough is central to its operation, as it ensures that computational resources are precisely allocated only where they are most impactful. By intelligently identifying and concentrating on the critical relationships within a vast token context, SubQ avoids the inefficiencies of processing irrelevant data points. Users have multiple avenues for engaging with SubQ: developers and enterprises can integrate it directly into their applications and existing enterprise workflows via its robust API, which supports a 12M token context window. Alternatively, for specialized use cases, SubQ Code can be seamlessly plugged into existing coding agents such as Claude Code, Codex, and Cursor, thereby providing them with an extraordinarily efficient and deep understanding of large codebases and extensive development histories.

Some common use cases include:

• Comprehensive Codebase Understanding: Enables AI systems and developers to reason across entire code repositories, facilitating deep understanding and interaction with vast amounts of source code for tasks like analysis, refactoring, and debugging.
• Persistent Agent State Management: Supports the maintenance of continuous and coherent historical context for AI agents over extended operational periods, leading to more consistent, informed, and intelligent interactions in complex scenarios.
• Advanced Software Engineering Tasks: Powers sophisticated real-world software engineering applications, as evidenced by its strong performance in rigorous benchmarks like SWE-Bench Verified, including automated bug fixing and code generation.
• Complex Linguistic Resolution: Handles intricate natural language processing tasks, such as multi-round coreference resolution, by accurately resolving references across numerous turns in lengthy conversational or document contexts.
• Streamlined Complex Software Workflows: Facilitates the execution of intricate processes that demand a profound and exhaustive understanding of extensive documentation, system logs, and detailed project histories, improving operational efficiency.
• In-depth Development Pipeline Analysis: Processes and analyzes months of pull requests and historical development data to extract critical insights, identify emerging trends, and provide comprehensive assistance for code reviews and project management.