Moonshot AI’s next flagship, Kimi K4, will have 5.4 trillion parameters and is expected in the second or third quarter of 2027, Chinese tech site MyDrivers reported on November 7, describing the roadmap as a leak. Moonshot has not confirmed it. By our arithmetic, 5.4 trillion is 1.93 times the 2.8 trillion of Kimi K3, the model Moonshot released in July.
Before K4, MyDrivers reported, Moonshot plans Kimi K3.5 at about 3 trillion parameters with 10 trillion post-training tokens, expected around the 2027 Lunar New Year, and a lighter Kimi K3 Flash this year to replace Kimi K2.8. ZOL (Zhongguancun Online) carried the same figures the same day.
Business Pill · Distillation explained
A Business Pill on distillation, how a small model learns from a large one. It is a general explainer and does not discuss Moonshot or Kimi.
The key insight: MyDrivers reported on November 7 that Moonshot plans Kimi K4 at 5.4 trillion parameters for the second or third quarter of 2027, and Kimi K3.5 at about 3 trillion with 10 trillion post-training tokens around the Lunar New Year. Moonshot has not confirmed it. By our arithmetic, K4 would be 1.93 times K3, and K3.5’s post-training tokens equal about 65% of K2’s 15.5 trillion pre-training tokens.
According to MyDrivers, Kimi K3 and Kimi K3.1 Pro both have 2.8 trillion parameters. K3.5 would rise slightly to 3 trillion, with its post-training data rising to 10 trillion, and is expected in 2027, possibly around the Lunar New Year. K4 would reach 5.4 trillion parameters, about twice K3, with release in the second quarter of 2027 or, later, the third.
MyDrivers said the K4 figure matches an earlier Nomura Securities estimate of 5 to 6 trillion parameters. It reported that Moonshot’s current computing power can support current research but has a limited ceiling, and that the focus is on self-distillation and post-training. ZOL reported that the team is concentrating on model compression, self-distillation and post-training optimisation, aiming to raise capability without significantly adding hardware.
Kimi K3 Flash is due within the year to replace Kimi K2.8, according to both outlets. MyDrivers said it was not known whether its performance and price could match DeepSeek V4.1 Flash.
Kimi total parameters, trillions: released K2 (1T), K3 (2.8T), K3.1 Pro (2.8T, per MyDrivers); reported, unconfirmed K3.5 (about 3T) and K4 (5.4T). Data: Moonshot AI; MyDrivers and ZOL, 7 Nov 2026, as reported. Our arithmetic: 5.4 / 2.8 = 1.93x.
What Was Posted on X
Today, leaked notes from a Moonshot investor call and Chinese tech reports laid out their full 2026 and 2027 lineup, and it explains how they plan to fix that problem before building their next giant model.
First, before the end of this year, Moonshot is releasing Kimi K3 Flash. That one is built to replace the older Kimi K2.8 and go head-to-head with DeepSeek V4.1 Flash.
Then around Q1 2027 (right around Lunar New Year),
they are shipping Kimi K3.5. Instead of making the base model bigger, they are keeping it near 3 trillion parameters and training it on 10 trillion post-training tokens. Their whole goal with K3.5 is self-distillation and compression basically teaching the model to reach the same answer using a fraction of the thinking tokens so user quotas actually last.
Once they get those running costs under control, the real monster comes in Q2 or Q3 2027: Kimi K4. According to the leak, Moonshot is scaling Kimi K4 all the way to 5.4 trillion parameters across their H100, H200, and B200 GPU clusters nearly double the size of K3.1 Pro.
That’s probably why Moonshot is looking to rent GPUs in Europe.
The account bebuu_ai circulated the roadmap in English on November 8, calling it notes from a Moonshot investor call plus Chinese tech reports. The K3 Flash, K3.5 and K4 figures match MyDrivers and ZOL; the investor-call framing, the quota goal and the H100, H200 and B200 cluster detail are the post’s own and are not in the Chinese reports we read.
On November 7, the account mylifcc posted the same roadmap in Chinese: K3.5 at about 3 trillion parameters in the first quarter of 2027 and K4 at 5.4 trillion in the second quarter, linking the timing to computing power and to the K3.1 Pro quota complaints.
Our Number: Post-Training at 65% of a Pre-Training Run
Moonshot’s Kimi K2 technical report says K2, a 1 trillion parameter model, “was pre-trained on 15.5 trillion tokens”. By our arithmetic, the 10 trillion post-training tokens reported for K3.5 equal about 65% of that entire K2 pre-training corpus. The K3 announcement on Moonshot’s blog gives K3’s size, 2.8T parameters, but we did not find a K3 pre-training token count there, so the comparison uses K2.
On the parameter side, our arithmetic puts K3.5’s reported 3 trillion at about 7% above K3’s 2.8 trillion, while K4’s reported 5.4 trillion is 1.93 times K3 and 5.4 times K2.
The K3.1 Pro Quota Episode
The leak follows a quota problem at launch. Moonshot released Kimi K3.1 Pro on October 24, and in the early hours of October 26 the Kimi assistant’s official account apologised for rate limits and abnormal quota consumption hitting subscribers, Cailian Press reported via MyDrivers. All active subscribers received a usage credit of 25% of their monthly total, according to that report.
The X post that circulated the leak in English, from the account bebuu_ai, says K3.5’s goal is self-distillation and compression so the model reaches the same answer with fewer thinking tokens “so user quotas actually last”. That framing is the account’s; the Chinese reports describe the focus on self-distillation and post-training without the quota link.
As we read the MyDrivers and ZOL reports, the near-term model, K3.5, adds little size and much more post-training, while the size jump is left to K4.
As we read the reports, the stated focus on self-distillation and post-training sits next to MyDrivers’ line that Moonshot’s computing power has a limited ceiling.
As we read the 26 October apology reported by Cailian Press, quota consumption on K3.1 Pro was already a subscriber issue before the roadmap surfaced.
bebuu_ai on X, 8 November 2026
“Their whole goal with K3.5 is self-distillation and compression basically teaching the model to reach the same answer using a fraction of the thinking tokens”
Three Implications
FOR MOONSHOT’S ROADMAP MyDrivers reports K3 Flash this year, K3.5 around the 2027 Lunar New Year and K4 in the second or third quarter of 2027; Moonshot has not confirmed any of them.
FOR POST-TRAINING By our arithmetic, K3.5’s reported 10 trillion post-training tokens equal about 65% of the 15.5 trillion tokens K2 was pre-trained on, per Moonshot’s K2 report.
FOR SUBSCRIBERS Cailian Press reported that Kimi gave all active subscribers a usage credit of 25% of their monthly total after the K3.1 Pro quota problems.
The essay says “Training builds the brain. Inference is the brain working.”
As we read it, the reported K3.5 plan spends on training, 10 trillion post-training tokens, to cut the thinking tokens paid at inference, per the X post’s framing. The essay does not discuss Moonshot or Kimi.
What Is Not Established
Moonshot has not announced Kimi K3 Flash, K3.5 or K4, and its official channels show no comment on the reported roadmap; its latest Hugging Face release is Kimi K3.1 Pro. MyDrivers describes the figures as a leak and does not name its source; ZOL reports them as Moonshot’s plans without naming a source. The English-language post calls them notes from a Moonshot investor call; the Chinese reports we read do not say that.
The post also says K4 will be trained across H100, H200 and B200 clusters, a detail we did not find in the Chinese reports.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
A roadmap reported by MyDrivers, which calls it a leak, and by ZOL puts Kimi K4 at 5.4 trillion parameters in the second or third quarter of 2027, 1.93 times K3 by our arithmetic, with a 3 trillion K3.5 trained on 10 trillion post-training tokens around the Lunar New Year. That post-training figure is about 65% of K2’s whole pre-training corpus, by our arithmetic. Moonshot has not confirmed any of it.
A note on sourcing. The K3 Flash, K3.5 and K4 details are from MyDrivers, which describes them as a leak, and ZOL (both 7 November); Moonshot has not confirmed them. The K3.1 Pro apology is from Cailian Press via MyDrivers (26 October). K2’s token count is from Moonshot’s K2 technical report. The 1.93x, 65% and 7% figures are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
