![\"Tencent](\"https://cms-image.pandaily.com/1/tencent_hy3_254f0e738f.png\")
Tencent Holdings reported unaudited second-quarter 2026 results on August 12, posting revenue of 204.8 billion yuan, with fintech and enterprise services at 60.3 billion yuan, up 9% year-over-year. On the earnings call, Chairman and CEO Pony Ma laid out a three-layer AI strategy: intelligence, applications, and infrastructure. The intelligence layer is anchored by the Hunyuan Hy3 model, released recently and rapidly integrated into products including Yuanbao. The model is now open globally through the API and Tencent's WorkBuddy international version. Hy4 is planned for near-term release.
The application layer showed the strongest traction. WorkBuddy, Tencent's AI-native office efficiency agent, posted rapid user growth with healthy retention across iOS, Windows, Android, and HarmonyOS. The company calls it China's leading AI-native office agent, citing engineering capability, model selection, and skills library. The portfolio has expanded from single-product strength to a matrix including Marvis and QClaw, and CodeBuddy achieved breakthrough user growth alongside WorkBuddy.
In the infrastructure layer, Tencent significantly increased compute procurement to convert model and application usage into revenue. Chief Strategy Officer James Mitchell said that given surging demand and rising lease prices, Tencent could almost immediately recover depreciation costs by renting compute to third parties, but chose a different path: directing most new compute toward training industry-leading models and deploying its own AI applications, believing superior intelligence converts to superior economics over the long run, including token sales through WorkBuddy.
President Martin Lau elaborated on capital expenditure. Tencent sees its business as two parts: the existing business, which generates stable growth with strong operating leverage, and the new AI-native business of self-developed models, new applications, and associated compute. The AI-native capex is an upfront investment for model training, future inference demand, and Tencent Cloud's AI capacity. Lau said compute ordered or prepaid months ago could now be resold at a margin above 30% over the original purchase price, evidence of how quickly AI compute economics have shifted.
Asked whether WorkBuddy is enterprise software or a new platform, Lau said it is a new platform: a flexible workspace for agentic AI aimed at all office productivity needs and one-person companies, with a harness layer leveraging different models and a growing library of third-party skills. Hy will be one of the models WorkBuddy offers, the primary one if it solves problems cost-effectively, but not the only one. Management framed the quarter as the beginning of an AI-empowered Tencent, with marketing services growing, several new games performing well, and WeChat Channels video views rising rapidly.
","publishedDate","2026-08-13T01:56:47.093Z","featured",false,"viewCount",0,"readingTime",3,"status","draft","originalWordPressId","featuredImage","featuredImageUrl","https://cms-image.pandaily.com/1/img_1_f7560054f6.png","gallery",[],[],"tags",[344,351,355],{"_316":345,"_11":346,"_319":347,"_348":349,"_34":350},"z98hhx0a0jb5fmtj001d79aj","Huawei","huawei","color","#6366f1","",{"_316":352,"_11":353,"_319":354,"_348":349,"_34":350},"g515globgbs9w747iwugdlvr","Global AI Giant","global-ai-giant",{"_316":356,"_11":357,"_319":358,"_348":349,"_34":350},"bgt1gjuk1a2h4yehlgdwbi1h","AI Research Center","ai-research-center","authors",[],"seo","externalVideo","createdAt","2026-08-13T01:56:31.893Z","updatedAt","2026-08-13T01:56:47.005Z","publishedAt","2026-08-13T01:56:47.054Z","featuredPosts",[371,400,427,458,488],{"_316":372,"_129":373,"_319":374,"_321":375,"_323":376,"_325":377,"_327":378,"_329":330,"_331":379,"_333":380,"_6":7,"_335":-5,"_336":-5,"_337":381,"_56":382,"_342":383,"_359":396,"_361":-5,"_362":-5,"_363":397,"_365":398,"_367":399},"wwn0lrk9krq502dtxmtqmfzr","DeepSeek V4 Pro Goes Live, Closes In on Frontier Tier","deepseek-v4-pro-goes-live-closes-in-on-frontier-tier","DeepSeek has rolled out DeepSeek-V4-Pro-0813 on its API platform, pairing a one-million-token context window with explicit agent-developer tooling and a 3x price premium over the Flash tier.","DeepSeek has pushed its V4 Pro model to general availability, listing the release as DeepSeek-V4-Pro-0813 on its pricing page. The flagship tier ships with a 1M-token context window and a 384K-token maximum output, defaulting to a thinking mode while still exposing a non-thinking endpoint for latency-sensitive calls.
\nFor long codebases, large document corpora and multi-step agent runs, the combination of 1M input and 384K output means a single call can ingest and generate far more material than the previous generation. DeepSeek has also wired in the capabilities most agent developers now expect: structured JSON output, tool calls, the Responses API and Anthropic-API compatibility, plus beta support for conversation-prefix continuation and FIM completion, with FIM restricted to the non-thinking mode. Wire-level, the service exposes both OpenAI- and Anthropic-compatible endpoints, so teams already running on either standard can switch without re-plumbing their stack.
\nPricing, meanwhile, reflects a deliberate tier split. Pro is set at 0.025 yuan per million tokens for cache hits, 3 yuan for cache misses on input and 6 yuan on output. Flash, by contrast, is priced at 0.02, 1 and 2 yuan respectively, putting Pro at roughly three times the cost of Flash on cache-miss input and output. Concurrency limits reinforce the divide: Flash supports 2,500 concurrent requests, while Pro is capped at 500.
\nThat ratio points to a clearer product structure for the V4 family. Flash is positioned for high-frequency, high-throughput traffic where cost dominates. Pro concentrates capacity on heavier reasoning, long-context coding and agent workflows where the model has more headroom to spend. The absolute price of Pro remains modest by frontier-model standards, leaving DeepSeek room to recoup the cost of the recent price-increase announcement while still undercutting most overseas competitors.
\nWith DeepSeek-V4-Pro-0813 now live, the V4 lineup is effectively a two-track offering: a low-cost, high-concurrency Flash tier for scale, and a higher-spec Pro tier aimed squarely at coding and agent use cases.
","2026-08-12T16:19:00.516Z",true,2,"published","https://cms-image.pandaily.com/1/img_ds_99a782bccc.png",[],[384,388,392],{"_316":385,"_11":386,"_319":387,"_348":349,"_34":350},"l42xhcsepyvfcsemx6tdd8p2","DeepSeek","deepseek",{"_316":389,"_11":390,"_319":391,"_348":349,"_34":-5},"oolipt2y8r0mi4x29gx8cs5m","DeepSeek V4","deepseek-v4",{"_316":393,"_11":394,"_319":395,"_348":349,"_34":-5},"e27atgrdlj2dfizg7mkjc2ld","AI Agent","ai-agent-1779329970",[],"2026-08-12T16:02:35.680Z","2026-08-12T16:19:00.436Z","2026-08-12T16:19:00.479Z",{"_316":401,"_129":402,"_319":403,"_321":404,"_323":405,"_325":406,"_327":378,"_329":330,"_331":332,"_333":380,"_6":7,"_335":-5,"_336":407,"_337":417,"_56":418,"_342":419,"_359":423,"_361":-5,"_362":-5,"_363":424,"_365":425,"_367":426},"ogjsldhrew0qtgbe1zv1gw2b","XPeng Opens Pre-Sale of G9L Flagship SUV at 25.98 Million Yuan — and Previews Second-Gen VLA 6.3.0","xpeng-g9l-flagship-suv-presale-25-98-million-yuan-vla-6-3-aug2026","XPeng unveiled the G9L flagship five-seat SUV in Guangzhou on August 11 with a 25.98 million yuan pre-sale price across four trims and dual powertrains. The launch also previewed the second-generation VLA model 6.3.0, which triples end-side model parameter count and adds L4-grade Robotaxi features including curbside parking and园区漫游.","Chinese electric vehicle maker XPeng unveiled the G9L flagship five-seat SUV in Guangzhou on August 11, opening pre-sale at 25.98 million yuan. The vehicle is offered in four trims and two powertrains, covering up to 755 kilometers of CLTC battery-only range and 1,602 kilometers of combined range. The launch also previewed XPeng's second-generation VLA driving model version 6.3.0, the first major upgrade since release.
\n\n\n\nThe G9L is the second pillar of XPeng's flagship dual-king strategy alongside the existing GX. The two share XPeng's SEPA 3.0 architecture and align on safety, intelligent driving, three-electric systems, handling, and comfort. The G9L has been validated for three years across 26 countries, accumulating more than 6.74 million test kilometers and satisfying over 3,000 regulatory items in 64 countries. The exterior is led by former Ferrari designer Juanma Lopez on a 5,120 by 1,999 by 1,782 millimeter body with a 3,100 millimeter wheelbase.
\n\n\n\nThe cabin introduces three globally-first technologies. The first is an all-body AI active-fit seat that detects body type and posture and adjusts geometry to match, rated to fit 98 percent of global body types. The second is an AI active-fit passenger-side zero-gravity seat with 24-way electric adjustment, integrated ABTS seatbelt and anti-submarine cushion airbag. The third is an intelligent four-motor executive console with electric voice opening and a 360-degree adjustable table. The vehicle adds a 270-degree sunshade, a 12.5-liter refrigerated console, a 21.4-inch 3K entertainment screen, a 33-speaker audio system, and 720-degree noise reduction.
\n\n\n\nThe intelligent driving upgrade is the more strategic piece of the announcement. VLA 6.3.0 lifts end-side model parameter count by a factor of 3.5 and perception sensitivity by 300 percent. The version expands cross-vehicle and cross-country generalization and adds VLA plus VLM cabin-driver fusion, with several L4-grade Robotaxi capabilities downshifted to consumer use, including start-from-stop, curbside parking, and campus-roam parking-spot search. The model also enables fuzzy-navigation, voice curbside parking, and voice-nearby parking through the VLM large model. The 6.3.0 release is positioned as XPeng's first major step toward L4-grade driving in mass-market consumer vehicles.
\n\n\n\nThe five-seat large-SUV segment in China is increasingly crowded. The G9L pitch is that differentiation is not just size but the integrated intelligent-driving and cabin stack. The pre-sale opens against competitive pressure from Li Auto, AITO, NIO's Onvo, and Xiaomi's YU7, all of which have launched five-seat flagship SUVs in 2026. The vehicle's market reception in the coming weeks will be the cleanest early read on whether XPeng's VLA-driven software story is strong enough to anchor a flagship launch.
","2026-08-12T02:26:28.354Z",{"_316":408,"_409":410,"_411":-5,"_412":-5,"_413":414,"_415":416},"wk7whhmfljwhdc9pz2lz0k4x","url","https://cms-image.pandaily.com/1/1_d3fe294eb3.png","alternativeText","caption","width",1920,"height",1080,"https://cms-image.pandaily.com/1/img_1_ad6882717a.png",[],[420,421,422],{"_316":345,"_11":346,"_319":347,"_348":349,"_34":350},{"_316":352,"_11":353,"_319":354,"_348":349,"_34":350},{"_316":356,"_11":357,"_319":358,"_348":349,"_34":350},[],"2026-08-12T02:05:03.358Z","2026-08-12T02:26:28.265Z","2026-08-12T02:26:28.320Z",{"_316":428,"_129":429,"_319":430,"_321":431,"_323":432,"_325":433,"_327":378,"_329":330,"_331":332,"_333":380,"_6":7,"_335":-5,"_336":434,"_337":-5,"_56":439,"_342":443,"_359":454,"_361":-5,"_362":-5,"_363":455,"_365":456,"_367":457},"p205d1xiom8bzdm8b0j6bqm5","Tencent Expands International Access to Flagship Hy3 Model, Offers Free WorkBuddy Access Until Aug 31","tencent-hy3-model-international-rollout-workbuddy-aug2026","Tencent rolls out its flagship Hy3 LLM internationally via WorkBuddy, Tencent Design Miora, and TokenHub — free on WorkBuddy until Aug 31, 2026.","HONG KONG, Aug 5, 2026 — Tencent today announced a broader international rollout of its flagship large language model, Hy3 (formerly known as Hunyuan), making it available to developers, users and enterprises through three entry points: the agentic workplace platform WorkBuddy, AI-native creative studio Tencent Design Miora, and Tencent Cloud's model platform TokenHub.
\n\n
Hy3 will be available free of charge to users worldwide on WorkBuddy until Aug 31, 2026 (Pacific Time), Tencent said, with further integrations planned across its AI-native product portfolio. Developers can connect to the model via API and integrate it with developer environments, coding extensions, third-party platforms and other tools.
\n\nOfficially released on July 6, Hy3 is built on a hybrid fast-and-slow-thinking mixture-of-experts (MoE) architecture with 295 billion total parameters and 21 billion active parameters, supporting a context length of up to 256K. Tencent said the model performs comparably to flagship models with two to five times as many parameters across reasoning, instruction understanding, in-context learning, code generation and agent capabilities. The model is available through international third-party platforms including OpenRouter, Hermes Agent, Kilo Code and Claude Code, as well as Hermes, Kilo, Cline, OpenClaw, OpenCode and Cherry Studio, and has been open-sourced under the Apache 2.0 license on Hugging Face, ModelScope, GitHub and GitCode.
\n\nThe South China Morning Post noted that the rollout is the first major update in Tencent's AI strategy since former OpenAI researcher Yao Shunyu joined the company to lead its foundational AI development, and that the international expansion is the latest move in a wave of Chinese technology firms taking aim at the world's leading AI players.
\n\nTencent said Hy3's API call volume has increased more than 68-fold compared with its previous-generation model, and that it topped OpenRouter's global LLM usage leaderboard within a week of launch. However, according to OpenRouter's weekly token usage ranking as of Aug 5, Hy3 currently ranks third, behind DeepSeek V4 Flash and Xiaomi MiMo-V2.5. API pricing on OpenRouter starts at US$0.1288 per million input tokens and US$0.5336 per million output tokens.
\n\n\"Hy3 marks a significant step in how Tencent Cloud International is helping organisations move beyond AI experimentation and put intelligence to work in practical, trusted and scalable ways,\" said Poshu Yeung, Senior Vice President of Tencent Cloud and Head of Tencent Cloud International. \"By connecting advanced models with real-world scenarios, enterprise-grade engineering, flexible model orchestration and secure infrastructure, we aim to give enterprises the control and reliability they need to move from pilots to scaled deployment.\"
","2026-08-05T08:16:37.000Z",{"_316":435,"_409":436,"_411":-5,"_412":-5,"_413":437,"_415":438},"zvot8j5y5axvfgy0aufm47fl","https://cms-image.pandaily.com/1/tencent_hy3_254f0e738f.png",1009,582,[440],{"_316":441,"_11":65,"_319":64,"_348":442,"_34":350},"ou7tkn9cq6eba9jr68hqiwch","#000000",[444,448,450],{"_316":445,"_11":446,"_319":447,"_348":349,"_34":350},"njwm066t5dgnkyahw9o23jvu","Tencent","tencent",{"_316":449,"_11":65,"_319":64,"_348":349,"_34":350},"ji5tef72gjtbd7bx0er6q9jv",{"_316":451,"_11":452,"_319":453,"_348":349,"_34":-5},"wbp115vvzeffpp45oh695q2z","Large Language Model","large-language-model",[],"2026-08-05T08:16:37.977Z","2026-08-05T08:48:28.960Z","2026-08-05T08:48:29.019Z",{"_316":459,"_129":460,"_319":461,"_321":462,"_323":463,"_325":464,"_327":378,"_329":330,"_331":379,"_333":380,"_6":7,"_335":-5,"_336":465,"_337":-5,"_56":470,"_342":475,"_359":484,"_361":-5,"_362":-5,"_363":485,"_365":486,"_367":487},"vm46u55wc5nh25vzl65yr8y6","World Third Strongest With Fable-Level Feel: Moonshot AI 2.8 Trillion Parameter Kimi K3 Breaks Open-Source Records","kimi-k3-2-trillion-open-source-record-jul2026","Moonshot AI releases Kimi K3 with 2.8 trillion parameters, ranking third globally behind Fable 5 and GPT-5.6 Sol, with open-weight release and API service.","Moonshot AI has officially released Kimi K3, a 2.8 trillion parameter open-source foundation model that ranks as the world third strongest on comprehensive benchmarks, with developer tests describing it as delivering Fable-level user experience. The model, already available with API service and developer documentation, is the largest open-source model by parameter count globally and represents a significant milestone in China AI development toward closing the gap with leading US foundation models.
Kimi K3 introduces Moonshot self-developed Kimi Delta Attention (KDA) mechanism combining linear attention and Attention Residuals (AttnRes) structure to optimize long-sequence information processing and deep network information flow. The model supports 100 million token context windows with up to 6.3x decoding acceleration through KDA, while AttnRes improves training efficiency by approximately 25% at less than 2% additional cost.
The model focuses on long-horizon agentic coding and self-evolving workflow capabilities. With a higher-sparsity MoE design featuring 896 expert modules while activating only 16 per inference, K3 balances expanded model capacity against controlled computational cost. It can understand large codebases, call tools, execute tests, and continuously adjust task plans based on feedback, representing a shift from answering questions to completing tasks.
Benchmark results place Kimi K3 in the global top tier. In the GDPval-AA v2 knowledge work test, K3 scored 1687 points, surpassing Claude Opus 4.8 Max (1600) and trailing only Claude Fable 5 Max and GPT-5.6 Sol Max. In the AA-Briefcase agent knowledge work test, K3 scored 1527 points, ranking second only to Claude Fable 5 Max. The model also achieved 91.2 on BrowseComp for long-context information retrieval, leveraging its 100 million token context window without requiring additional context compression.
Developer community feedback highlights K3 autonomous planning and execution capabilities in complex agent tasks, continuous programming, and multi-tool orchestration. Several hedge fund evaluations rated K3 performance above Opus 4.8 at approximately 60% of its pricing. The open-weight release strategy follows DeepSeek precedent, allowing global developers to freely download, deploy, and modify the model, creating competitive pressure on closed-model strategy companies like OpenAI and Anthropic. At a time when Anthropic plans to raise Opus 4.8 prices by 50% in September, K3 and other Chinese open-source models present a cost-effective alternative that could reshape the global AI model market dynamics.
","2026-07-17T08:00:00.000Z",{"_316":466,"_409":467,"_411":-5,"_412":-5,"_413":468,"_415":469},"ygx4v5pu8cgxic843zqrtthq","https://cms-image.pandaily.com/1/article1_e6769801a4.png",1672,941,[471],{"_316":472,"_11":473,"_319":474,"_348":442,"_34":-5},"nd7u5jlcbw44fqibmzu4ckhw","Technology","technology",[476,480,481],{"_316":477,"_11":478,"_319":479,"_348":349,"_34":350},"q5ktywk4z70rf8zxnq5bbabt","Moonshot AI","moonshot-ai",{"_316":449,"_11":65,"_319":64,"_348":349,"_34":350},{"_316":482,"_11":483,"_319":483,"_348":349,"_34":350},"cipqssdhm3pbuuk4gple3qxz","kimi",[],"2026-07-17T02:02:16.538Z","2026-07-17T03:38:01.739Z","2026-07-17T03:38:01.811Z",{"_316":489,"_129":490,"_319":491,"_321":492,"_323":493,"_325":464,"_327":378,"_329":330,"_331":379,"_333":380,"_6":7,"_335":-5,"_336":494,"_337":-5,"_56":497,"_342":500,"_359":513,"_361":-5,"_362":-5,"_363":514,"_365":515,"_367":516},"im82va13re6zj59a49b2ul13","WAIC 2026 Top 10 Highlights: Chinese Chips Line Up in Full Force as AI Enters the Industrial Era","waic-2026-top-10-highlights-chinese-chips-jul2026","WAIC 2026 opens with 100+ chip companies, supernode clusters, humanoid robots, and 64 consumer AI products across three Shanghai zones.","The 2026 World Artificial Intelligence Conference opened in Shanghai on July 17 across three major zones covering Zhangjiang, World Expo, and Xuhui West Bund, forming the largest WAIC in history. The theme Intelligent Partners, Co-Creating the Future reflects a fundamental shift from model capability demonstrations to industrial application deployment, with domestic chip full-lineup showcases, supernode computing clusters taking center stage, humanoid robots entering real factories, and consumer AI products becoming hands-on experiences.
Zhangjiang Science Hall presents the hardware ceiling of the conference with over 100 computing chip and infrastructure companies displaying 200 exhibits including 67 domestic first launches. A dozen general computing chip companies compete on the same stage. Highlight exhibits include Dongfang Suanxin DF1000, China first large-scale AI chip built on mature 14nm process claiming 4nm-equivalent performance through a third-way architecture innovation. Other chip companies including Muxi, Tianshu Zhixin, and Moore Threads showcase their latest products with live training demonstrations.
The World Expo H2 hall is dominated by supernode systems. Huawei debuts Atlas 950 SuperPoD with 1024 Ascend cards on display, while the Atlas 950 SuperCluster can scale to 500,000 cards. Sugon Dawn 8000 scaleX shows the 100K-card fully domestic AI super-fusion cluster with over 300 optimized applications. Alibaba Zhenwu M890 and ZTE OEX supernodes complete the lineup, enabling direct comparison of competing supernode technical approaches.
H3 Embodied Intelligence Hall features 161 companies and 314 exhibits. Xuanshi highlights include Unitree GD01 world first manned deformable mecha capable of switching between humanoid and quadruped forms. Zhiyuan demonstrates a production line replicated on the expo floor. Fourier debuts GR Nano desktop companion robot. Qianshun Intelligence Moz1 performs long-horizon single-instruction room organization tasks. Multiple companies demonstrate robot pharmacies, coffee bars, and other real-world applications.
Sustainable Xuhui West Bund presents 64 consumer AI products for hands-on experience. Honor Robot Phone, Bubbo 1 companion robot, multiple AR smart glasses from Rokid, Liangliang Vision, Xinmou Technology, and Ant Group are available for user testing. Music generation, video creation, and virtual human interaction stations round out the consumer experience zone. The conference structure itself signals that AI industry has moved beyond model competition into a phase where infrastructure, terminals, applications, and ecosystems compete as integrated systems.
",{"_316":495,"_409":496,"_411":-5,"_412":-5,"_413":468,"_415":469},"nap5nzjw3ij04ndv27sgqv4s","https://cms-image.pandaily.com/1/article2_fb9350fec2.png",[498],{"_316":499,"_11":59,"_319":58,"_348":442,"_34":350},"u9uewvwvdfl4hd8yy6xyqkwq",[501,505,509],{"_316":502,"_11":503,"_319":504,"_348":349,"_34":350},"pn2of6p1b3ukwko2147hx8rm","WAIC","waic",{"_316":506,"_11":507,"_319":508,"_348":349,"_34":-5},"kntxzq3cpf7q91yqp2bvpe1h","AI Infrastructure","ai-infrastructure-tag",{"_316":510,"_11":511,"_319":512,"_348":349,"_34":350},"x4rcbmu97kiw4ayqunjpq2lb","AI Chip","ai-chip",[],"2026-07-17T02:02:33.894Z","2026-07-17T03:37:24.641Z","2026-07-17T03:37:24.746Z","isDraft"]