A new AI lab calling itself Subquadratic launched on 5 May 2026 with $29 million in seed funding and a product that breaks one of deep learning's deepest assumptions. Its first model, SubQ 1M-Preview, is not a transformer. Instead of standard O(n²) transformer attention, SubQ uses sparse subquadratic attention end-to-end, and ships with a native context window of 12 million tokens — roughly twelve times larger than the current frontier.

Why subquadratic matters

Transformer attention scales quadratically with context length. Doubling the input doubles the compute per attention head twice. Practical models hit a wall around 1M–2M tokens because the memory bandwidth and the matrix-multiplication cost both blow up. Subquadratic architectures — Mamba, RWKV, RetNet, and now SubQ — replace the dense attention matrix with sparser operations whose compute grows much more slowly with context. The historical trade-off was quality. Subquadratic models lagged transformers of a similar size on benchmarks. SubQ's interesting claim is that it has closed this performance gap.

The vendor claims

Per WhatLLM's reporting, SubQ claims:

  • Roughly 1/5 the cost of frontier transformer models on long-context tasks (think: full-book summarisation, multi-file code analysis, regulatory document review)
  • Up to 52× faster attention at scale when context exceeds a few million tokens
  • Native 12M-token context without resorting to retrieval-augmented sliding windows or context-compression hacks

These are the company's own numbers. Independent benchmarks have not yet been published, and as Air Street Press observed in its May state-of-AI roundup, the gap between vendor claims and HELM-Lite or LongBench scores on subquadratic models has historically been wide.

What 12M tokens unlocks

At 12M tokens you can fit roughly nine novels, or the source code of a mid-sized open-source project, or a full year of board minutes and quarterly filings, in a single prompt. Use cases that require reasoning over a whole corpus, not just retrieving from it, become viable. Finding every contradiction across a year of board minutes is unreachable for RAG but should be possible if the model's attention is truly subquadratic.

For anyone building LLM-backed finance or legal workflows, the practical signal is that long-context AI is moving from niche experimentation to a credible commercial offering. The next twelve months will tell whether SubQ's benchmarks survive contact with independent evaluation. If they do, the transformer-or-nothing assumption that has dominated the field since 2017 is about to be tested for the first time at production scale.

Independent evaluation arrived early, and the vendor paid for it

The wait for independent evaluation was expected to be a year; it took three months. The result, however, comes with a significant qualification.

Appen published a third-party assessment of Subquadratic's preview models across long-context retrieval, code generation, business-workflow automation and graduate-level reasoning. Subquadratic engaged Appen to carry it out. A paid-for benchmark is a normal commercial arrangement, but it is not the same as an unaffiliated lab testing a competitor's claims — the kind of result needed to overcome historical skepticism.

On the axis tested, the numbers held. Sparse self-attention measured 56 times faster than FlashAttention-2 at one million tokens, against the 52-fold figure the company claimed. Retrieval was effectively perfect at the one-million and two-million tiers, returning the target value verbatim on every sample, and the nano variant held 98 per cent exact-match at six million and twelve million tokens.

The axis that historically broke these architectures is still unpublished

Subquadratic models have not usually failed on speed or on retrieval. They have failed on quality against transformers of comparable parameter count, which is why the closed-gap claim was the interesting one.

The published benchmark set is narrow: three tests, all of them long-context retrieval or coding. Official figures include 95.0 per cent on RULER at 128K, 65.9 per cent on MRCR v2 at one million tokens, and 81.8 per cent on SWE-Bench Verified. Nothing has been published on general reasoning, mathematics, multilingual performance or safety.

The existing benchmarks measure retrieval at twelve million tokens, which is a different capability from reasoning over that same context. Finding a value verbatim in nine novels demonstrates that attention reaches the whole context. It does not demonstrate that the model can hold a contradiction between month three and month eleven of a set of board minutes, which was the use case that made the architecture interesting in the first place.

What a buyer can now conclude

Long-context work that is fundamentally retrieval — find the clause, locate the function, pull the figure — has a credible commercial option that did not exist in May, and the efficiency claim behind it has been checked by someone other than the company, even if the company commissioned the check.

Long-context work that is fundamentally reasoning is still unevaluated. Anyone building a finance or legal workflow on this should be clear about which of the two they are actually buying, because the published evidence separates them cleanly and the marketing does not.

The transformer-or-nothing assumption has been tested at production scale on one dimension — retrieval — and passed. The other dimensions are what made it an assumption in the first place.

Sources and cross-checks: Primary: WhatLLM — New AI Models May 2026. Corroborated against: Air Street Press — State of AI: May 2026. Funding amount ($29M), context window (12M tokens), and "1/5 cost / 52× faster" claims verified as vendor-stated; awaiting independent benchmarks. Cross-checked 18 May 2026. Retrospective added 23 August 2026: the Appen third-party assessment (commissioned by Subquadratic), the 56x FlashAttention-2 comparison, the retrieval results at the 1M-12M tiers and the RULER, MRCR v2 and SWE-Bench Verified figures are as reported by Appen and DataCamp, 2026. No general-reasoning, mathematics, multilingual or safety evaluations had been published as of that date.