K3发布后的八大要点：模型、云计算与应用的连接

China (PRC) I Internet 7 Takeaways Post K3 Release: Connect TheDots on Models, Cloud &Applications intelligence. ARR has fastest growth the next day; (2) More T-parametersChina models (e.g. Qwen3.8, 2.4T) to come; (3) Updates from Zhipu andStepfun; (4) Compute is focus on recent financing /KC Preview (link);(5)Token Expenditure Index remains soft;(6)Video modelslandscapedo not see notable change;(7) Tencent's WorkBuddy gains traction; (8)Reaffirm Buyon BABA's full stack capabilities pricing comparison between China and Us, model intelligence from Artificial Analysis, tokenconsumption in OpenRouter, Token Expenditure Index trends and other industry data foranalysis. intelligence milestone and ranked at top positions in different metrics. According to Arena(Exhibit 5).This is the first time China takes the lead over US models in Frontend Code Arenavs early 2025 (DeepSeek-R1 is the last Chinese models came close with US models). Investorsfeedback is more about the closing of intelligence gap and anticipate more Chinese models(Exhibit 2) having T-parameter in coming months. Balance between intelligence and good touse amid supply constraints For Chinese models is expected. Moonshot Al: Still far from the upper limit in model intelligence. ARR reached the fastestScaling Lawisvalid.K3 isnot themost intelligentmodeldespiteabig advancement is madevsK2.6; (2) It is still early to talk about the upper bound of modelintelligence; (3) ARR experiencedthefastestgrowthratethenextdaypostK3release. Latest strategies from Zhipu and Stepfun on industry trends.(1) Zhipu: (a) Scaling Lawcontinues to accelerate across data, environment, RL and infra. (b) GLM-5.2 has lessparameters vs K3.It can exceed K3 capabilities when scale to latter's level.Please refer topage 2for details;(2) Stepfun: During WAIC (a)agents become the atomic unit of productivity:(b) integration between Al and terminals is the next focus; (c) Al coding exceeds human; (d)Agent-to-Agent is the next internet Compute is the focus on recent financing by Al Lab and Kingsoft Cloud Preview. RecentMiniMax(UsD2bn,Jul-26).Weviewnewfundingcan strengthen respective computing needsamid surging demand. For Kingsoft Cloud, we adjust our revenue slightly in 2Q, factoring inthe supply constraints issues with capex spending happened more in Jun vs Apr/May (link) Token expenditure trends reflect highly cost efficient model is important. According toSilicon Data LLM Token Expenditure Index (Exhibit 4), the index continues to remain soft andranged between 1.56 to 1.61the week of 12 Jul (vs 1.63 to 1.65 the week of 5 Jul), vs 2.04 on31 May. Based on Artificial Analysis, China models are highly cost-efficient vs US models andAPI cost is a fraction vs US models (Exhibit 3) . We look into (1) Blended price vs Intelligenceby models (Exhibit 1).The recent industry trends from Us /China tech companies show thedebate on Tokenmaxxingvs token spending means cost efficientmodels become increasingimportant in Coding and Agentic Al. ThomasChong*IEquityAnalyst852374380161thomas.chong@jefferies.com Video generationmodel landscapedoes not seeto have notablechange.According to Kimi'sco-founder Mr. Zhou Xinyu, the current focus is model intelligence and video generation model ZoeyZong*Equity Analyst85237438163|zoey.zong@jefferies.com Jefferies releasedSeedance2.5andconsidersvideo Jefferies others (link) continues to gain traction. According to Analysys (link), it highlights operating metrics forWorkBuddy in the segment (Mar-26). These include (1) number of desktop monthly visits reached8.85m; (2) WorkBuddy is ahead of Trae (ByteDance), QClaw (Tencent) and Qoder Work (BABA)On the other hand, BABA releases Meoo (diverse Al enterprise apps) during WAIC. For Baidu, ithighlights a number of Al showcases such as Baidu Drive, Baidu Wenku and Xiaodu during WAiC(link). On the other hand, Kingsoft Office held Al Productivity Conference last week and released Alproducts Lingxi Professional and WPS Al Hub (link) What's next to be anticipated ? These include (1) With model iterations happen in less than 3months, we expect upcoming models for Qwen (Qwen 3.8, 2.4T), MiniMax (M3 Pro), Zhipu (GLM)and Tencent (Hy4) are to be anticipated. For BABA, it highlights Qwen3.8 is open weight model (totalin Token Plan, Qoder and Qoder Work. It is also available in PC and mobile of Qianwen; (2) a numberof companies are expected to achieve over USD1bn ARR in 2026 or early 2027:(a) BABA; (b) Zhipu,(c) MiniMax; (d) Kling; (e) Volcano Engine. We expect Kimi ARR is to be watched out post releaseof K3. We view BABA to stand out on its full stack of capabilities with models, cloud, chips andapplications Zhipu:KeyTakeawaysfromConferenceCall Scaling Law continues to accelerate. This can be seen in breakthrough made by OpenAl andAnthropic heading towards AGlThoughts on K3. (1) K3 is successfulin scaling and close to global frontier models; (2) GLM-5.2has lessparameters vsK3 (Exhibit6).Itexpectsto exceedK3 capabilities