AI

Subquadratic Launches SubQ, an LLM Built on a Subquadratic Architecture

San Francisco-based AI company Subquadratic has released SubQ 1M-Preview, a large language model (LLM) that the company states is built on a fully subquadratic architecture. Unlike traditional transformer models, where computational requirements scale quadratically with context length, the company claims SubQ’s compute grows linearly, potentially enabling significantly longer context windows and lower operational costs.

The announcement, made by co-founder and CEO Justin Dangel, marks the public introduction of a model designed to address a fundamental scaling limitation in AI. The company has raised $29 million in seed funding from investors including Justin Mateen, co-founder of Tinder, and Grant Gittlin of Lasagna, alongside early backers of organizations such as Anthropic and OpenAI.

Addressing the Quadratic Scaling Bottleneck

The core technical challenge the company aims to solve is the quadratic scaling of traditional transformer models. In standard architectures, every token must be compared against every other token within a given context. As context length grows, the number of these interactions—and the required compute—increases quadratically. Dangel noted that this constraint has historically shaped system design, forcing developers to use workarounds such as retrieval-augmented generation (RAG) pipelines and chunking strategies rather than processing full document corpora directly.

“Developers and investors spend more of their time and money on workarounds than on the problem itself,” Dangel stated in the announcement. “RAG systems use a search engine to pull a small number of relevant results before sending them to the model, because sending the full corpus isn’t feasible.”

Subquadratic’s new architecture is designed from the ground up to be subquadratic, a goal that the company acknowledges the research community has pursued for years through mechanisms like linear attention and state space models. The breakthrough, according to the company, lies in achieving this efficiency without sacrificing the performance level of existing frontier models.

Product and Performance Benchmarks

SubQ 1M-Preview is being made available through a private beta, offering three initial products:

  • API: A full-context API for developers and enterprise teams.
  • SubQ Code: A coding agent that loads entire codebases into a single context window, accessible via a command line interface (CLI). This allows for planning, execution, and review across a full repository in one pass.
  • SubQ Search: A long-context search tool designed to provide deep research capabilities at chatbot-like speed.

The company published several performance benchmarks for SubQ 1M-Preview, which it said were verified by a third party on the RULER 128K benchmark, a standard test for reasoning over extended inputs. Key results include:

  • A score of 95% accuracy on the RULER 128K benchmark, compared to 94.8% for Claude Opus 4.6.
  • The company’s Sparse Attention mechanism was reported to be 52 times faster than FlashAttention in its architecture-level comparison, while requiring 63% less compute.
  • On the MRCR v2 benchmark (which tests retrieval and reasoning across a long context), a third-party verified score of 65.9 was reported. This compares favorably with scores cited for GPT 5.5 (74), Claude Opus 4.7 (32.2), and Gemini 3.1 Pro (26.3).
  • A SWE-Bench Verified score of 81.8, compared to Opus 4.6 (80.8) and Deepseek 4.0 Pro (80.0).

The company also shared a research result demonstrating performance on up to 12 million tokens, noting that other frontier models “break down well before their stated 1M-token limit.” At this scale, SubQ’s architecture reduces attention compute by almost 1,000x compared to other frontier models, according to the company.

Enabling Long-Context Applications

By reducing the cost and compute associated with processing large amounts of data, SubQ is positioned to enable new categories of applications. The company suggests that expanding the frontier for context window size makes it viable to process entire codebases, large document collections, complex spreadsheets, or long-running interaction histories in a single pass. This approach reduces the need for complex retrieval pipelines and multi-agent coordination systems, and helps preserve information that might otherwise be lost at context boundaries.

Economics and Team

Cost is highlighted as a primary constraint in deploying AI systems at scale. By making inference more efficient, Subquadratic aims to make workloads that were previously cost-prohibitive viable. This includes high-volume workloads and applications that rely on sustained, long-context interaction with the model.

The company’s team includes 11 PhD researchers and research engineers with backgrounds from organizations such as Meta, Google, Oxford, Cambridge, and Adobe. CTO Alex Whedon previously worked as a software engineer at Meta and served as Head of Generative AI at TribeAI, where he led over 40 enterprise AI implementations.

Conclusion

With the release of SubQ 1M-Preview, Subquadratic is entering a competitive market for large language models, offering a fundamentally different architectural approach. The company’s claim of efficient linear scaling with context length, if validated at scale, could shift the practical limits on what is possible with AI. As the company moves from private beta to broader availability, the technology’s performance in real-world, high-volume applications will determine its long-term impact.


Source: https://subq.ai/introducing-subq