美银中国互联网AI观察KimiK3发布我们的初步看法202607175页

AI Watch:Kimi K3 launch–our initialthought Industry Overview Welcome to the latest“AI Watch”series in which we look at China’s rapidly evolving AIlandscape and implications for China Internet sector. 17 July 2026 EquityChinaInternet/e-Commerce Kimi K3lifts open-source frontier to near-Fable 5 level MoonshotAI (private) launched Kimi K3 model on July 16, a 2.8-trillion-parameter MoEmodel (16 of 896 experts activated) with a 1M-token context window and nativemultimodal input—the largest open-weight model released globally to date, with fullopen-weight on July 27. Architecturally, K3 debuts Kimi Delta Attention and AttentionResiduals technology, driving improving capability with efficient cost structure. Onperformance, K3 took #1 on Arena's Frontend Code leaderboard, a 17-place jump overK2.6, ahead of Claude Fable 5 and GPT-5.6 Sol (see page 2 for benchmark details). Onpricing, Moonshot AI priced K3 at a premium $3/$15 USD per million input/outputtokens—the most expensive Chinese model to date. K3 pricing is at 60% of ClaudeOpus 4.8/around half of GPT-5.6 Sol level, and a sharp step up from K2.6 at $0.95/$4USD. Comparing to Chinese SOTA, K3 is priced at 2-3x of Zhipu’s GLM5.2. Alex Liu>>Research AnalystMerrill Lynch (Hong Kong)+852 3508 5090alex.liu3@bofa.com Miranda Zhuang, CFA>>Research AnalystMerrill Lynch (Hong Kong)miranda.zhuang@bofa.com Joyce Ju>>Research AnalystMerrill Lynch (Hong Kong)joyce.ju@bofa.com Joanna Du>>Research AnalystMerrill Lynch (Hong Kong)joanna.du@bofa.com Who isMoonshot AI (Kimi)? Kimi is the flagshipAI model series developed by Moonshot AI, a Beijing-based leadingChinese independent AI lab founded in March 2023 by Tsinghua alumni Yang Zhilin, ZhouXinyu and Wu Yuxin. The company built its reputation on long-context processing andpivoted decisively to open-source with the 1T-parameter Kimi K2 in July 2025, followedby K2 Thinking and the multimodal K2.5/K2.6 series. Cursor's Composer 2, a popularglobal coding model, was later revealed to be built on K2.5 base. Per Bloomberg, Alibabaand Tencent both participated in previous fund-raising of Moonshot AI. SOTA: state of the art GPU: Graphics Processing Unit MoE: Mixture of Experts Implications to ChinaAI/Internet sector and key stocks Our high-level takeaway is, despite persistent hardware/compute capacity constraints inChina, K3 demonstrates that pre-training scaling, paired with architectural innovation,can still deliver step-change gains for flagship Chinese models. Moonshot nearly tripledparameter scale vs. K2 (2.8T vs. 1T) and we note key Chinese model labs are allattempting to scale up in coming generations. For Tencent, we see K3 launch as neutralas Tencent is still catching up and thus benefits from the democratization of AI technicalknow-how driven by other AI labs. For Alibaba, while it benefits from broad AItraining/usage growth for its cloud service given tight compute environment, AlibabaQwen's "open-source leader" narrative may face some tests. Finally, K3 raises thecapability ceiling for China AI models, shifting the burden of proof to other independentAI labs(e.g. Zhipu and MiniMax).It also deepens investor skepticism about model-layermoats, as competitive leadership has become increasingly transient. Released coding benchmark comparison of Kimi K3 Analyst Certification We, Alex Liu and Joyce Ju, hereby certify that the views each of us has expressed in thisresearch report accurately reflect each of our respective personal views about thesubject securities and issuers. We also certify that no part of our respectivecompensation was, is, or will be, directly or indirectly, related to the specificrecommendations or view expressed in this research report. Disclosures Important Disclosures FUNDAMENTAL EQUITY OPINION KEY: Opinions include a Volatility Risk Rating, an Investment Rating and an Income Rating. VOLATILITY RISK RATINGS, indicatorsof potential price fluctuation, are: A - Low, B - Medium and C - High. INVESTMENT RATINGS reflect the analyst’s assessment of both a stock’s absolute total returnpotential as well as its attractiveness for investment relative to other stocks within its Coverage Cluster (defined below). Our investment ratings are: 1 - Buy stocksare expected to have a total return of at least 10% and are the most attractive stocks in the coverage cluster; 2 - Neutral stocks are expected to remain flat orincrease in value and are less attractive than Buy rated stocks and 3 - Underperform stocks are the least attractive stocks in a coverage cluster. An investment ratingof 6 (No Rating) indicates that a stock is no longer trading on the basis of fundamentals. Analysts assign investment ratings considering, among other things, the 0-12 month total return expectation for a stock and the firm’s guidelines for ratings dispersions (shown in the table below). The current price objective for a stockshould be referenced to better understand the total return expectation at any given time. The price obj