AI系列第55期：混元3正式版发布，Token趋势，锁仓到期

China (PRC) | Internet AI Series #55: Hy3 Official Version Release;Token Trends; Lock-Up Expiry Highlights (1)Tencent released Hy3 with further enhancements (e.g.success rate, cost efficiencies, long horizon tasks). API price is competitive(Exhibit 4) to support more users and scenario; (2)Token ExpenditureIndex: The week of 28 Jun remains soft and behind May level; (3)OpenRouter: Token usage is stable vs prior week; (4) Lock-up expiry forZhipu / MiniMaxthis week; (5) new measures aim to protect users from AIvirtual intimate relationship. In this report, we provide investors with over 50 data points on token consumption in OpenRouter,pricing for different models, model intelligence in Artificial Analysis,user trends by sub-sectorand other industry data for analysis. Hy3 official version has a number of enhancements following user feedback.Highlightsof Hy3 vs preview version include (1) feedback received across over 50 business units andupgrade with better user experience. On input/output price, Hy3 is charged at a lower price atRMB1/RMB4 per million tokens vs RMB2/RMB8 for preview version. It aims to support moreusers and scenario; (2) further enhancements in post training such as Reinforcement Learning(RL) Exhibit 5. Better performance in long contexts vs large models which are 2-5x more interms of parameters; (3) value creation in productivity scenario such as software development,office, financial, design and game development; (4) Blind test among 270 experts in real workscenario show Hy3 scores better than GLM-5.1 (e.g. frontend, data, storage, CI/CD); (4) furtherimprovements in (a) higher task success rate vs preview version (90% vs prior 72%), (b) tokenefficiency: compared to GLM-5.2, savings in token consumption is achieved in word processing(47.4%) and PPT (49%); (c) reduced hallucination by 15 pp. Please refer to Exhibit 6 acrossdifferent third party assessments. Silicon Data LLM Token Expenditure Index: The week of 28 Jun showed continued softness.The index (Exhibit 2) is weighted by where usage is concentrated (e.g. expenditure goes upwhen demand shifts towards premium models). On the other hand, media reports (e.g. TheInformation, Economic Observer) highlight tech and internet companies set upper limit ontoken consumption among employees. According to our checks, the index ranged from 1.62to 1.71 between 28 Jun to 4 Jul (vs 1.68 to 1.74 between 21 to 27 Jun), vs 2.04 on 31 May.Based on Artificial Analysis, China models DeepSeek, Qwen, Kimi, MiniMax and Zhipu are highlycost-efficient vs US models and API cost is a fraction vs US models (Exhibit 3), thanks toformer’s refined architecture such as MoE (Mixture of Experts), attention mechanism (Group-Query Attention GQA, sparse attention) and MLU (Model FLOPs Utilization). OpenRouter: Token consumption remains stable vs prior week.Based on analysis of theweek of 29 Jun (vs week of 22 Jun) (Exhibit 1), number of weekly tokens consumed remainedstable at 46.7tn. In terms of ranking, DeepSeek V4 Flash ranked no 1, followed by Xiaomi MiMo-V2.5 and MiniMax M3. DS-V4 and Seedance 2.5 will be released in Jul.As per our recent updates (link) and (link),DS-V4 will have new pricing strategies starting from 15 Jul with the doubling of input/outputprice during peak hours (9:00 to 12:00, 14:00 to 18:00) (Exhibit 7-9). We view highly cost-efficient model is key amid rising AI demand and supply constraints. On the other hand, videogeneration model Seedance 2.5 is expected to be released soon with upgrade in (a) 30-secvideo generation; (b) up to 50 multimodal input assets; (c) upgrades in video editing andgeneration improve efficiencies and quality. Thomas Chong * | Equity Analyst852 3743 8016 | thomas.chong@jefferies.com Zoey Zong * | Equity Analyst852 3743 8163 | zoey.zong@jefferies.com Continued overleaf... InternetEquity ResearchJuly 6, 2026 More products from AliCloud and Tencent Cloud will not be offered for free starting from 1 Aug.AliCloud:Databridge Agent (DTS) will start commercialization starting from 1 Aug. It connects withvarious mainstream database, perform data analysis, security and integration with AI ecosystem;Tencent Cloud: it will stop providing free quota for services such as AI native GPU Developmentand Deployment starting from 1 Aug. Meituan launches open-source LongCat 2.0 model.Highlights (1) total parameters 1.6T. Dynamicactivations between 33B to 56B tokens; (2) 1M context; (3) highly cost efficient with LSA (LongCatSparse attention LSA). It is highly token-level dynamic on compute through Zero-computationexperts + ScMoE; (4) use of domestic compute on agent, reasoning and interactions; (5) It is the firstmodel of this scale to complete full training and inference on a 50K-card domestic compute cluster;(6) high rankings vs peers in SWE-bench Pro (software engineering) and Terminal-Bench (agent). WorkBuddy takes the lead in Workspace Agents market.According to third party Analysys, ithighlights operating metrics for WorkBuddy i