Subquadratic's LLM claims face scrutiny
Subquadratic says its SubQ model runs far faster than standard attention, but researchers say the evidence is thin. A reminder to test AI claims yourself.
A Miami startup says it has cracked one of the most expensive problems in AI, and the claim is now drawing pushback. Subquadratic launched its SubQ model on May 5 with figures that sound too good to be true: by its own numbers the model runs about 52 times faster than FlashAttention while using 63% less compute, and cuts attention compute by nearly 1,000 times at a 12-million-token context.
The idea is what the company calls subquadratic sparse attention. Standard attention scales quadratically, so doubling the input roughly quadruples the work. SubQ instead routes each token only to the tokens that actually matter. Subquadratic reports a production model with a 1-million-token window, 95.6% on RULER at 128K (just ahead of Claude Opus 4.6 at 94.8%) and 81.8 on SWE-Bench Verified.
Why the doubt matters
On June 19, MIT Technology Review reported harder scrutiny than the launch suggested. A third-party evaluator told the publication the architecture could be a real advance, while other researchers said the public evidence does not yet justify the strongest claims and that limited outside access has made the results hard to reproduce. One comparison in the piece reached for the word Theranos.
The point for an operator is not whether SubQ is real. It is that a single vendor’s benchmark table, however striking, is a marketing artifact until someone outside the company reproduces it. Leaderboard scores routinely overstate what a model does on your real work.
What to do with a claim like this
Treat eye-catching numbers as a reason to test, not a reason to switch. Run the model on your own tasks before you trust it, and keep a verification habit for what it produces. The models worth adopting are the ones you can verify, not the ones with the loudest chart.