After the K3 release, the intelligence gap in Chinese AI models continues to narrow, and BABA's full-stack capabilities are again viewed favorably

Overview

This report focuses on Chinese AI models, cloud services, and the application ecosystem after the K3 release, providing data points on model pricing, China-U.S. price comparisons, the Artificial Analysis intelligence index, OpenRouter token consumption, and the Token Expenditure Index. The report argues that Chinese models are continuing to approach leading U.S. frontier models in intelligence, while also holding meaningful advantages in API cost and model efficiency; on the application side, enterprise AI and agent tools are beginning to show traction.

Core views

The core views include: first, K3 has reached a new intelligence milestone, but Moonshot AI believes model intelligence is still far from its upper bound; second, more China T-scale parameter models are about to launch, including Qwen 3.8, MiniMax M3 Pro, Zhipu GLM, and Tencent Hy4; third, recent AI Lab financing and Kingsoft Cloud's outlook show that compute remains the focus; fourth, the weakening Token Expenditure Index reinforces the importance of high cost-performance models in Coding and Agentic AI; fifth, no clear change has yet emerged in the competitive landscape for video generation models; sixth, Tencent WorkBuddy has gained traction on the application side; seventh, BABA is again viewed favorably for its full-stack capabilities across models, cloud, chips, and applications.

Analysis framework

The report uses a cross-validation approach based on industry data, combining model intelligence indices, model parameter scale, API input/output pricing, token consumption trends, company earnings calls, and industry conference information to assess the gap between Chinese and U.S. AI models, cost efficiency, compute demand, and the progress of application commercialization.

Methodology notes

• Model evaluationArtificial Analysis Intelligence Index v4.1

  Comprehensive intelligence index across nine benchmarks

  This index incorporates GDPval-AA v2, t-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR to compare the overall intelligence performance of different AI models.

• Pricing and efficiency analysisBlended Price vs Intelligence

  Comparison of model intelligence versus blended API pricing

  The report combines model intelligence levels with input, output, and cache pricing to assess the cost-performance and commercial usability of Chinese models relative to U.S. models.

• Demand trackingSilicon Data LLM Token Expenditure Index

  Token spending index

  This index is used to observe LLM token spending trends; the report notes that the index range for the week of July 12 was 1.56 to 1.61, below 2.04 on May 31.

• Peer comparisonFrontend Code Arena and Arena Elo

  Frontend coding capability and Arena Elo ranking

  The report cites Arena and Stanford AI Index data to show that the gap between Chinese and U.S. models is narrowing on metrics such as frontend coding.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Alibaba Group Holding Limited (BABA)

  The report reiterates BUY and argues that it has full-stack capabilities across models, cloud, chips, and applications.

  Strengths

  Qwen model iteration, open-source weights, and cloud and application distribution capabilities create a combined advantage.

  Weaknesses

  The report does not provide a specific target price or upside, and near-term commercialization still needs to be monitored through ARR and application adoption.

  Comparison

  Compared with pure-play model companies, BABA is seen as more prominent in full-stack capabilities.

  Risks

  Intensifying model competition, changes in token costs, and cloud demand falling short of expectations.

• Kingsoft Cloud (KC)

  The report mentions adjusting 2Q revenue expectations due to supply constraints and the June capex cadence.

  Strengths

  Growth in AI compute and cloud demand could create revenue opportunities.

  Weaknesses

  Supply constraints pressure near-term revenue recognition.

  Comparison

  Similar to the beneficiary logic from AI Lab financing, KC is more directly exposed to cloud infrastructure supply capacity.

  Risks

  Capex timing, supply chain constraints, and customer demand volatility.

• Tencent Holdings Ltd. (700 HK)

  The report focuses on Tencent WorkBuddy traction and mentions the upcoming Hy4 model.

  Strengths

  WorkBuddy desktop monthly visits reached 8.85m, with application-side user metrics leading some peer tools.

  Weaknesses

  The report does not provide full revenue conversion data.

  Comparison

  WorkBuddy is described as leading products such as Trae, QClaw, and Qoder Work.

  Risks

  Enterprise AI application conversion rates, competitive catch-up, and model update cadence.

• Moonshot AI / Kimi K3

  The report treats K3 as the core event behind the model intelligence breakthrough and the change in industry sentiment.

  Strengths

  K3 is close to global frontier models, achieved the fastest next-day ARR growth after release, and reached 2.8T in parameter scale.

  Weaknesses

  Its average speed of 20 TpS is below the average 60 TpS shown for GLM-5.2 in the table.

  Comparison

  GLM-5.2 has fewer parameters and lower pricing, but K3 stands out more in scale and multimodal context.

  Risks

  Subsequent model iteration, cost control, and the competitive landscape after open-source weight release.

• Zhipu GLM

  The report summarizes Zhipu's earnings call views, emphasizing the Scaling Law and GLM-5.2 capabilities.

  Strengths

  GLM-5.2 offers 1M context, lower API pricing, and higher average speed, while emphasizing infrastructure efficiency.

  Weaknesses

  Its current parameter scale is below K3.

  Comparison

  If scaled to K3's size, Zhipu believes GLM-5.2 could surpass K3.

  Risks

  Scale expansion, training cost, and execution of RL and infrastructure optimization.

Key data

• Kimi K3 total parameters2.8TThe report table lists Kimi-3 total parameters at 2.8T.
• DeepSeek-V4-Pro total parameters1.6TThe report lists DeepSeek-V4-Pro total parameters at 1.6T.
• Qwen3.7Max total parameters1.2TThe report lists Qwen3.7Max total parameters at 1.2T and mentions that the Qwen 3.8, 2.4T model is about to be released.
• GLM-5.2 total parameters744BGLM-5.2 has fewer parameters than K3, but Zhipu believes it could surpass K3 if scaled to K3's size.
• MiniMax M3 total parameters428BThe report lists MiniMax M3 total parameters at 428B and mentions that MiniMax M3 Pro is worth watching.
• Hy3 total parameters295BThe report lists Hy3 total parameters at 295B and mentions that Tencent Hy4 is worth following later.
• GLM-5.2 API pricingRMB2/8/28；USD0.26/1.4/4.4The table uses cache/input/output pricing.
• Kimi K3 API pricingRMB2/20/100；USD0.3/3.0/15.0The table uses cache/input/output pricing.
• Token Expenditure Index1.56-1.61Range for the week of July 12, below 1.63-1.65 for the week of July 5 and 2.04 on May 31.
• WorkBuddy desktop monthly visits8.85mAnalysys data, for March 2026.
• Recent AI Lab financingDeepSeek USD7bn；Zhipu USD4bn；MiniMax USD2bnThe report says the financing could strengthen related companies' demand for compute.

Impact & implications

The investment implication of the report is that Chinese AI models are improving simultaneously in intelligence and cost efficiency, which could enhance the competitiveness of Chinese internet platforms, cloud vendors, and AI application companies. Compute constraints and capital expenditure remain important variables for converting cloud revenue, while progress in application-side products such as WorkBuddy, Meoo, Baidu Drive, Baidu Wenku, Xiaodu, and WPS AI Hub will determine whether AI capabilities can be translated into ARR and user growth.

Risks

• Compute supply and capex timing may constrain cloud revenue realization.
• A persistently soft Token Expenditure Index may reflect insufficient demand or spending intensity.
• Rapidly changing competition between Chinese and U.S. models means a narrowing intelligence gap does not imply long-term leadership.
• No significant change has been seen in the video generation model landscape in the short term, and related commercialization priority may remain below model intelligence.
• Enterprise AI application traction still needs to convert into ARR and paying customers, and some current metrics are still more operational than financial.
• Jefferies discloses investment banking or service relationships with some covered companies, and investors should note potential conflicts of interest.

What to watch

• The sustainability of ARR after the Kimi K3 release and the impact of open-source weight release.
• Subsequent model launches such as Qwen 3.8, MiniMax M3 Pro, Zhipu GLM, and Tencent Hy4.
• BABA's progress in coordination across Qwen, cloud, chips, and enterprise applications.
• Easing Kingsoft Cloud supply constraints and the timing of capex deployment.
• Activity levels and revenue conversion of enterprise AI applications such as WorkBuddy, Meoo, Baidu AI applications, and WPS AI Hub.
• Whether the Token Expenditure Index and OpenRouter token consumption resume growth.
• Changes in API pricing, intelligence indices, and Coding/Agent task costs for Chinese models relative to U.S. models.