SubQ 1.1 Version Released Amid Community Skepticism

ME News reports that on June 17 (UTC+8), according to monitoring by Beating, SubQ—a controversial large model previously claiming to reduce computational consumption by a thousandfold—has released a technical report for its 1.1 Small (small-parameter) version. In response to earlier criticisms that the preview version lacked a paper or independent validation and was mocked by the community as an “AI Swiss Army knife” (implying false advertising), developer company Subquadratic partnered with evaluation firm Appen to conduct a third-party assessment, claiming the model achieved 98% retrieval accuracy at a maximum length of 12 million tokens and performed close to state-of-the-art mainstream models in practical programming tests. The technical report also revealed that the model was not trained from scratch but was instead modified by replacing the attention computation mechanism in an open-source state-of-the-art model and incrementally training it on one trillion tokens. Even with third-party validation, the developer community remains skeptical of this update. Some researchers argue that the so-called “black tech” offers no fundamental breakthrough and merely implements an existing technique—block-sparse attention—by splitting long texts into chunks and dynamically filtering them. Others criticize the report for including AI-generated filler text, particularly evident in section 5.7.1. System engineers warn that the filtering mechanism introduces additional scheduling overhead under concurrent multi-user usage, causing severe lag for the slowest 1% of users. Since the model’s core parameters have not been made publicly available for download, nor has a publicly accessible API been opened, the promised reductions in computational demand and ultra-low pricing remain purely theoretical. (Source: BlockBeats)