A more attentive AI could cost significantly less. With SubQ, the American startup Subquadratic promises to reduce the computational burden in large language models. The approach seems viable, but the published figures still require thorough verification.
Why SubQ Targets the Economic Weak Point of Language Models
Large language models still heavily rely on the Transformer, the architecture popularized in 2017. Its central mechanism, attention, compares each piece of text with all others. Like in a room where everyone speaks to everyone else, the clamor quickly grows overwhelming.
This cost follows a quadratic logic: doubling the text does not merely double the workload; it multiplies it by four. For a company analyzing contracts, customer tickets, or source code, this rule can swiftly turn a long request into a hefty computing bill.
What Subquadratic Really Announces with SubQ 1M-Preview
Subquadratic, co-founded and led by Justin Dangel, unveiled SubQ on May 5, 2026. The company claims that its model utilizes sparse attention, meaning it only considers a subset of the relationships between tokens.
A token represents a fragment of a word or sign that the model processes. SubQ aims to identify useful connections rather than comparing all positions. The promise lies in a clear concept: maintain the valuable conversations, cut the rest, and preserve coherence.
The figures presented set a high standard. Subquadratic mentions handling 12 million tokens in a search result, achieving 52 times more speed than FlashAttention at 1 million tokens, and nearly 1,000 times less attention computation in certain scenarios.
Why Developers View This Announcement with Caution
Long context windows intrigue teams working with code repositories, document databases, or regulatory files. Without complete context, they break down files, index passages, and use RAG, or retrieval augmented generation, which involves searching before generating.
SubQ promises to simplify these scaffolds, but a closed model is difficult to evaluate. The published benchmarks, such as RULER or SWE-Bench Verified, provide reference points. They cannot replace independent audits on varied corpora, assessing costs, errors, and reproducible limits.
The Real Issue Lies More in Verified Uses than the Announcement
The idea of lighter attention is not new with SubQ. Researchers have been exploring linear, sparse, or recurrent approaches for several years. Recent theoretical work also shows that certain forms of rapid attention can remain feasible, but under precise mathematical conditions.
The central question then becomes operational. SubQ must demonstrate that a long context not only swallows a lot of text but also retrieves the right phrase at the right moment. Having an entire book in memory means little if the useful page disappears.
Subquadratic is launching SubQ via an early access API, SubQ Code for software repositories, and SubQ Search for long-range searching. The next benchmark will be straightforward: public tests, conducted repeatedly by third parties, on 1 million tokens and beyond.



