Mocked as "AI catch-all cure" by critics, SubQ releases version 1.1: invites third-party evaluations to prove itself, but is accused of incorporating AI boilerplate language

According to Beating monitoring, the controversial large model SubQ, which previously claimed it could reduce computational consumption by a thousand times, has released a technical report for its 1.1 Small (small-parameter) version.

In response to accusations that earlier preview versions were mocked by the community as an “AI panacea” (meaning false advertising) due to the lack of papers and independent verification, the R&D company Subquadratic, together with evaluation firm Appen, conducted a three-party assessment. It claims that the model achieved 98% retrieval accuracy under a maximum sequence length of 12 million tokens, and also performed at a level close to mainstream leading models in hands-on programming tests. The technical report also discloses that the model was not trained from scratch; instead, based on an open-source cutting-edge model, it was adapted by replacing the attention computation mechanism and then incrementally trained for 1 trillion tokens.

Even after bringing in three-party evaluations to prove it, the developer community still remains full of doubts about this update. One researcher noted that the so-called “black technology” actually does not involve any underlying technical breakthroughs; at its core, it is simply an existing technique that cuts long text into small blocks and then performs dynamic filtering (i.e., block-sparse attention mechanisms). Some readers also complained that the technical report included AI-generated filler phrases (especially evident in section 5.7.1). System engineers, meanwhile, warned that when the filtering mechanism is used concurrently by multiple people, it would introduce additional scheduling overhead, causing severe lag for the slowest 1% of users.

Because the model neither publicly releases the core parameters for everyone to download nor opens up an API interface that everyone can use, the so-called promises of reduced compute and ultra-low pricing still remain nothing more than paper claims for now.