Found first: a primary source the press has not covered yet.
A paper posted to arXiv on August 28 finds that Sliding Window Attention with sinks matches or outperforms post-trained linear attention on every benchmark tested, with margins of 2 to 10 times on long-context reasoning tasks. The authors close with an explicit recommendation: practitioners should switch to SWA instead of post-training linear models.
What the source says
The paper is by Alexia Jolicoeur-Martineau, Rhea Sanjay Sukthanker, Pashmina Cameron, and Emy Gervais; institutional affiliations are not listed on the abstract page. The authors compare SWA with sinks against post-trained linear attention across multiple benchmarks, including Needle-in-a-Haystack and BABILong. On long-context reasoning tasks, SWA scores 2 to 10 times higher. They describe SWA as requiring no post-training, "extremely fast," and "low memory," and state that linear attention would need to be trained from scratch or given extensive post-training to reach comparable performance.
Why it matters
Multiple labs are currently spending post-training compute converting pretrained models to linear attention, on the premise that linear attention is more memory-efficient at long contexts. This paper tests that premise directly on the long-context benchmarks that motivate the conversion, and finds SWA wins by a substantial margin with no post-training cost at all. The recommendation in the paper is not hedged: switch. For teams currently making architecture decisions around long-context inference, the paper presents SWA as both cheaper to deploy and stronger on the tasks that matter.