Dubbed the "AI Jack-of-All-Trades," SubQ has released version 1.1: While bringing in third-party evaluations for self-confirmation, it has been accused of incorporating AI buzzwords.

According to Dynamic Insight monitoring, the controversial large model SubQ, which claimed to reduce computational costs by a thousandfold, has released a 1.1 Small Parameter version technical report.

In response to early criticisms of the preview version for lacking a paper and independent validation, the R&D company Subquadratic conducted a third-party evaluation in collaboration with the assessment firm Appen. They claimed that the model achieved 98% retrieval accuracy under a maximum length of 12 million tokens and performed close to state-of-the-art models in real-world programming tests. The technical report also revealed that the model was not trained from scratch but was based on an open-source cutting-edge model. It replaced the attention calculation mechanism and underwent incremental training with 1 trillion tokens.

Despite the third-party validation, the developer community remains skeptical of this update. Some researchers pointed out that the so-called black technology did not actually have a fundamental technological breakthrough. Essentially, it only applied existing technology, such as block-sparse attention mechanisms, by segmenting long texts into small blocks for dynamic filtering. Some readers also criticized the presence of AI-generated text clichés in the technical report (particularly evident in section 5.7.1). System engineers warned that the filtering mechanism would introduce additional scheduling overhead in multi-user concurrent use, resulting in severe latency for the slowest 1% of users.

Since the model has neither publicly shared core parameters for download nor provided an open API interface for everyone to use, the promised computational efficiency reduction and ultra-low pricing commitment remain merely theoretical.