Navigating China Internet AI models Focuses on harness & model performance vs

Equity Research 3 August 2026 | 6:57PM HKT

NAVIGATING CHINA INTERNET/AI MODELS

Focuses on harness & model performance vs. price; multi-modal highlights; top AI/apps tracker

Ronald Keung, CFA | Goldman Sachs (Asia) L.L.C.

Lincoln Kong, CFA | Goldman Sachs (Asia) L.L.C.

From cost efficiency, rise in model intelligence to new frontier/increasing pricing Timothy Zhao | power of Chinese AI models (our LLM primer), we now estimate Chinese AI model Goldman Sachs (Asia) L.L.C. Annualized Recurring Revenues (ARRs), in aggregate, to reach US$13bn by year-end 2026 (prior: US$10bn) on higher demand for Chinese models/faster ARR ramp up Steve Qiu | with our lifted Zhipu API/MiniMax group year-end ARR at US$2.5bn/US$1bn. We Goldman Sachs (Asia) L.L.C.

expect competition for the best performance-to-price balance to intensify amongst Eunice Liu | Chinese AI models, evident from MiniMax’s H3’s open-weight approach priced at Goldman Sachs (Asia) L.L.C.

~1/3 of peers’ pricing, and DeepSeek’s latest V4 Flash (at 284bn parameters, Luqing Zhou | reaching close to GLM5.2 in front-end coding) and Alibaba Qwen3.8 Max (at 2.4T Goldman Sachs (Asia) L.L.C.

parameters) official launches, both reached new performance-per-dollar efficient Damian Xie | frontiers (as per Arena.ai for front-end coding). We believe Zhipu’s GLM-5.5 and Goldman Sachs (Asia) L.L.C. MiniMax’s M3 next update/M3 Pro models will be next to watch. Iris Xiao | Goldman Sachs (Asia) L.L.C. We highlight the latest developments around 1) multi-modal: Launch of MiniMax’s H3 on July 31 where full modal-capabilities, open-weight approach and significantly lower pricing/costs are key highlights, while Bytedance’s Seedance 2.5 (API channel launch on July 31) continues to lead in long-duration generation (30s max length) and consistency (50 multi-modal reference inputs); expanding agentic functionalities will drive further content creation adoption/penetration in our view; 2) our anticipation of rise in competition of higher-end coding/agentic-focused models, at the 2-5 trillion parameter size class over the next few months as scaling continues, where Alibaba Qwen3.8 Max launched at 2.4T total parameter size moved up in Arena.ai’s overall Code Arena and front-end coding rankings; 3) coding/agentic performance uplift in the smaller-sized/agentic-focused segment, with DeepSeek V4 Flash official API launch, outperforming its earlier V4 Pro Preview; and Tencent’s Hy3 marked performance uplift vs. Hy3 Preview. We believe post-training around supervised fine-tuning (SFT) has driven the key performance enhancements; 4) harness/agentic applications, where internet mega-caps are leading in workspace applications, Tencent’s Workbuddy and Alibaba’s product revamp with QwenWork; and 5) continued focus on AI capex

Goldman Sachs does and seeks to do business with companies covered in its research reports. As a result, investors should be aware that the firm may have a conflict of interest that could affect the objectivity of this report. Investors should consider this report as only a single factor in making their investment decision. For Reg AC certification and other important disclosures, see the Disclosure Appendix, or go to Analysts employed by non-US affiliates are not registered/qualified as research analysts with FINRA in the U.S.

Goldman Sachs Navigating China Internet/AI models

spending, on the back of anticipated increasing domestic compute supply over 2H26. Factoring in the latest industry landscape and ARR developments, we raise Z.AI and MiniMax 2026E revenues by 35%/63% respectively, fine-tune Z.AI/increase MiniMax 2027-28E estimates and maintain Neutral on Zhipu, Buy on MiniMax on risk-reward.

Estimate changes for Z.AI and MiniMax: Factoring in latest industry landscape and ARR developments, for Z.AI, we change our 2026-28E revenues by +35%/+4%/-2% and our earnings estimates by -3%/-3%/-1% on strong 2026 YTD ARR ramp up, while factoring in intensified competition in longer term over larger parameter/high-end coding models and sustained R&D investments. For MiniMax, we raise our 2026-28E revenue by 22-64% on favorable mix shift towards fast growing API business from AI-native products, and fine-tune our…