Kimi Researcher Challenges NaiveAI Claim That 2,000 tok/s Shows "AI Building AI"

NaiveAI, founded by Tsinghua University associate professor Dai Jifeng, has released its first model, N0.5-Flash, and promoted it with the claim that "AI participated in developing AI." The company said the model was involved in architecture design, code writing, and inference optimization, while reporting a peak single-stream inference speed of 2,122 tok/s. That framing was quickly challenged by Yang Xinyu, a researcher at Moonshot AI, the company behind Kimi. Yang said reaching 2,000 tok/s is entirely feasible on its own, but argued that NaiveAI used a weak comparison baseline. He noted that the company compared its result against SGLang at a little over 300 tok/s, while Xiaomi had previously pushed the larger MiMo-V2.5-Pro past 1,000 tok/s. Yang also disputed NaiveAI's characterization of its architectural changes. He said N0.5-Flash was modified from Xiaomi's MiMo-V2.5 Base by replacing global attention with DSA and ultimately using a DSA + GQA setup. In his view, that looks more like an engineering trade-off built on an existing base model, one that may also reduce GPU utilization, rather than a major architecture breakthrough. He added that NaiveAI has shown a model can be made faster through continued training and engineering optimization, but has not shown that "AI participation in R&D" produced a breakthrough beyond what human teams can achieve.