Kimi researcher publicly questions NaiveAI’s claims about “AI participating in research and development AI” and its speed breakthrough
BIBIBI
AT A GLANCE
NaiveAI claims N0.5-Flash achieved 2122 tok/s and was developed by AI participated in the R&D, but the related publicity has been questioned.
Article
A Kimi researcher questions NaiveAI, 2000tok/s unable to prove “AI AI AI”
Dongcha Beating AI breaking news: NaiveAI, founded by Tsinghua University Associate Professor Dai Jifeng, has just released its first model N0.5-Flash. The company promotes “AI participation in R&D AI” and says the model participated in architecture design, code writing, and inference optimization. The officially announced peak single-stream inference rate reached 2122 tok/s.
Yang Xinyu, a researcher at Moonshot AI, subsequently publicly questioned this publicity. He believes that 2000 tok/s is entirely possible in itself, but NaiveAI chose a baseline that was too weak. It used SGLang’s 300-plus-tok/s as the reference; Xiaomi had previously already run the larger-parameter MiMo-V2.5-Pro at over 1000 tok/s. With a low baseline, the final improvement naturally appears more striking.
Yang Xinyu also does not accept NaiveAI’s description of its architectural innovation.N0.5-Flash is based on Xiaomi’s MiMo-V2.5 Base, replacing the original global attention with DSA. NaiveAI ultimately used DSA + GQA. Yang Xinyu believes this is more an engineering trade-off based on an existing base model, and it also sacrifices GPU utilization, so it cannot be considered a major architectural breakthrough.
He believes NaiveAI has demonstrated that a model can become faster through continued training and engineering optimization, but it has not yet proved that “AI participation in R&D” itself has produced a breakthrough beyond what a human team could achieve.
Original link https://m.theblockbeats.info/flash/369342
Key points
01
Tsinghua University Associate Professor Dai Jifeng founded NaiveAI.
02
NaiveAI released its first model N0.5-Flash, claiming that AI participated in architecture design, code development, and inference optimization.
03
NaiveAI announced a peak single-stream inference rate of 2122 tok/s。
04
Yang Xinyu, a Moonshot AI researcher, believes that 2000 tok/s itself is entirely capable of achieving this, but NaiveAI adopted a relatively weak comparison baseline.
05
NaiveAI used SGLang's 300 more tok/s as a reference; Xiaomi previously had the larger-parameter MiMo-V2.5-Pro reach 1000 tok/s or more.
06
N0.5-Flash Based on Xiaomi's MiMo-V2.5 Base, NaiveAI ultimately adopted DSA+GQA。
07
Yang Xinyu believes this is more of an engineering trade-off based on an existing base model, possibly sacrificing GPU utilization, and cannot be considered a major architectural breakthrough.
08
Yang Xinyu believes NaiveAI has proven that a model can become faster through continued training and engineering optimization, but has not yet proven that AI participating in the R&D led to breakthroughs that human teams could not achieve.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
NaiveAI showcased a very high inference speed and claimed that it AI participated in the model's R&D. Yang Xinyu's questions focused on two aspects: whether the speed comparison was fair, and whether the architectural changes were sufficient to be called a major innovation.
Why it matters to readers
This dispute shows that model peak speed needs to be assessed in conjunction with a standardized benchmark, hardware configuration, and independent verification; looking at a single speed figure cannot prove AI that participation in the R&D produced a unique breakthrough.
tok/s refers to the number of tokens generated per second. A higher figure generally indicates faster generation, but figures from different test conditions cannot be compared directly.
Risks and unknowns
The evidence comes from a BlockBeats relay on Telegram and does not provide an official technical report, independent reproduction, or a complete link to the original statement.
The evidence does not specify the 2122 tok/s corresponding hardware, concurrency settings, or context length.
The evidence does not provide a unified baseline or third-party verification; therefore, the speed and architecture assessments cannot be independently verified on this basis.
The evidence does not quantify AI what contributions participation in research and development made that human teams could not.
Related Developments
Loading event timeline…
Related concepts
tok/s
This term is not in the glossary yet. Browse related concepts in the glossary.
