【速報】Kimi K3がオープンウェイト化。2.8兆パラメータの中身を完全解剖

The weights and technical report for Kimi K3 were released on July 27, 2026. The technical report is 47 pages long, includes a technical blog, and the model itself is 1.56TB. It's rare for all of this information to be made available at once. However, I doubt many people will read the entire document in Japanese. Therefore, I've reviewed everything and extracted only the most interesting parts. The stated specifications (2.78 trillion parameters / approximately 100 billion movements per iteration / 93 layers) I recounted the weights myself using the publicly released weight index and confirmed that they all match. ■ Chapters 0:00 Hook 0:19 What was released? 0:54 What are parameters? 2:19 Who is the opponent? 3:28 It can't be moved 4:43 Only the index was opened 5:25 Isn't that cutting corners? 6:43 There were 60 types of parts. 7:56 Mathematical formulas were included as physical objects. 9:23 Traces of human hands. 10:57 The top-ranked entry is eliminated due to certain conditions. 11:59 The company blog is too honest. 12:49 It can't be measured because it refuses. 15:02 Started making its own tools. 17:12 The last line of the acknowledgments. 17:56 What is open for? 19:10 Summary ■ Primary Source Technical Report: https://github.com/MoonshotAI/Kimi-K3... Technical Blog: https://www.kimi.com/blog/kimi-k3 Model Body: https://huggingface.co/moonshotai/Kim... Joint Evaluation by AISI (UK) and NIST CAISI (National Institute of Science and Technology): https://www.aisi.gov.uk/blog/prelimin... ■ Method Used for Verification The weights themselves (1.56TB) were not downloaded. The safetensors format contains a tensor index (name, type, shape) at the beginning of each file. Only the beginning portion was retrieved using an HTTP Range request, and the data for 96 files was aggregated. The total retrieved data was 72MB = 0.005% of the total. ■ Disclaimer ・The statement in the video, "If you let the AI ​​read everything, it would take 3.98 million times, about 1 year and 3 months, and about 1.9 billion yen," is an estimate based on assumptions. This is an estimate based on writing each weight into a 6-character text and converting it to 4 characters = 1 token. Since weights are not text in the first place, this method of reading them is not a practical application. The required number of GPUs (60 for RTX 4090 / 18 for H100) is an estimate based on the weight size. Actual inference requires additional memory such as KV cache, and the officially recommended configuration is 64 accelerators or more. The issue of "disappearing under the condition of being ranked 1st" in BrowseComp means that when K3 is measured with 1M full context, it scores 90.4, which is the same value as the GPT-5.6 Sol figure in the report. Since the measurement conditions for GPT are not publicly available, it cannot be said that they scored the same under the same conditions. The only comparison target for security evaluation is GLM-5.2. Claude and GPT were not measured because they refused the task, and therefore it cannot be said that "K3 is stronger." The chip design is at the RTL simulation stage (the report itself explicitly states "initial proof of concept"), and no actual device has been manufactured. The actual measured number of active parameters was approximately 105.8 billion. The difference from the nominal 104.2B appears to be due to whether or not embedded layers are included in the count, but this cannot be confirmed. *All figures are based on publicly available information as of July 28, 2026. ■ Credits Narration: VOICEVOX: Zundamon / VOICEVOX: Shikoku Metan BGM: Maoudamashii https://maou.audio/ Website: https://www.ai-jitan-hub.com