Kimi researcher questions NaiveAI, stating that 2,000 tokens per second does not prove "AI builds AI".

Beating AI News (Dongcha) – NaiveAI, founded by Tsinghua University associate professor Dai Jifeng, has just released its first model, N0.5-Flash. The company’s core proposition is “AI develops AI”, claiming the model participated in architecture design, code writing, and inference optimization. The official announced peak single-stream inference speed reaches 2122 tokens per second (tok/s). Yang Xinyu, a researcher at Moonshot AI, subsequently publicly questioned this claim. He noted that a speed of 2000 tok/s is achievable, but NaiveAI chose an overly weak baseline for comparison: the company used SGLang’s ~300 tok/s as a reference, while Xiaomi previously ran its larger-parameter MiMo-V2.5-Pro at over 1000 tok/s. A lower baseline makes the resulting improvement appear more dramatic. Yang also rejected NaiveAI’s description of its architectural innovation. N0.5-Flash is modified from Xiaomi’s MiMo-V2.5 Base, replacing the original global attention with DSA. NaiveAI ultimately adopted DSA + GQA. Yang argued that this is more of an engineering trade-off based on an existing base model, which also reduces GPU utilization, and does not qualify as a significant architectural breakthrough. He stated that NaiveAI has proven models can be accelerated via continued training and engineering optimizations, but has not yet demonstrated that “AI developing AI” itself delivers breakthroughs beyond what human teams can achieve.