Daily Report on Hong Kong-Listed LLM Concept Stocks (July 7, 2026)

I. Industry Observations

On July 7, 2026, Hong Kong stocks pared early gains to close lower. The Hang Seng Index fell 0.51% to 23,496.89 points, the Hang Seng Tech Index dropped 0.75% to 4,507.04 points, and the Hang Seng China Enterprises Index slipped 0.54% to 7,770.26 points. Daily turnover reached HK$319.711 billion, with southbound capital recording a net purchase of HK$497 million.

Performance across the AI and large model industry chains was mixed. Tencent Holdings rose 2.0% following the official release of its next-generation large model, Hunyuan Hy3, and a share repurchase worth HK$205 million. Meituan-W climbed 4.5% as LongCat-2.0 was officially open-sourced and successfully adapted by Moore Threads. Zhipu grew 5.2% upon completing its A-share IPO tutoring phase and recently launching the ZCode programming agent. On the other hand, Kuaishou-W plummeted 12.0%, and SenseTime-W slid 6.7%.

The semiconductor and hardware sectors experienced a sharp correction. Memory chip concepts led the decline, with GigaDevice plunging 12.2%, Montage Technology diving 10.2%, and SMIC dropping 3.0%. The PCB and optical communications sectors were also under pressure, as BroadEx Technologies fell 9.1%, Yangtze Optical Fibre and Cable decreased by 7.8%, and Kingboard Laminates retreated 5.2%. Embodied AI concepts remained active, with QuantGroup skyrocketing 55.7% after completing multiple rounds of technical verification centered around restaurant back-kitchen scenarios.

II. Market Performance of LLM Concept Stocks

Data Source: HKEX. For reference only, not investment advice. Data as of market close onJuly 7, 2026. 

III. Industry and Corporate Dynamics

1.Ant Lingbo: Releases Spatial Perception Model LingBot-Depth 2.0, Simutanously Open-Sourcing Vision Foundation Model LingBot-Vision

On July 7, Lingbo Technology, an embodied AI company under Ant Group, officially released its spatial perception model, LingBot-Depth 2.0, and simultaneously introduced its vision foundation model, LingBot-Vision. Trained on a dataset of 150 million records, LingBot-Depth 2.0 secured first place in 12 out of 16 evaluation benchmarks for depth completion. In scenarios with large-area missing depth data indoors, its depth error was reduced by approximately 50% compared to the previous generation, showing outstanding performance in environments where traditional depth cameras typically fail, such as glass, mirrors, and transparent objects. LingBot-Vision is the industry's first vision model to adopt "boundary structure" as its pre-training objective, enabling sub-pixel level boundary localization and spatial structure understanding. This open-source release includes model weights for four versions—ViT-G/L/B/S—alongside the technical report. Ant Lingbo also announced a deeper partnership with Orbbec, planning to launch an SDK integrated with LingBot-Depth capabilities and a commercial all-in-one camera.

2.Qianwen Large Model: Upgrades Real-Time Speech Recognition Model Fun-ASR-Realtime, Achieving Hundred-Millisecond First-Character Latency and Supporting 16 Dialects and 30 Languages

On July 6, Qianwen Large Model announced an upgrade to its real-time speech recognition large model, Fun-ASR-Realtime, controlling the latency of the first character within the hundred-millisecond range. Its streaming recognition accuracy is close to the level of offline models, and it supports 16 dialects and 30 languages. In speech recognition tests covering 16 dialects across eight major dialect regions, the model achieved an average character accuracy rate of 88.62%, leading in 12 dialect categories and surpassing related products from Volcengine and Tencent. The model has undergone specialized optimization for East Asian and Southeast Asian multilingual scenarios, such as Thai, boosting recognition accuracy by 20%, making it suitable for scenarios like overseas customer service and international conferences. The API has gone live on the Alibaba Cloud Bailian platform, and the offline mode model, Fun-ASR-Flash, was also launched recently.

3.OpenRouter: Weekly Invocation Volume of Chinese Large Models Leads the World for Ten Consecutive Weeks, with Chinese Models Securing the Top Six Spots

According to estimates based on the latest data from OpenRouter, the total global invocation volume of large AI models reached 46.7 trillion Tokens last week (June 29 to July 5). Among them, the weekly invocation volume of Chinese large AI models reached 23.45 trillion Tokens, representing a quarter-on-quarter increase of 15.01%. This marks the sixth consecutive week of growth and the tenth consecutive week that China has exceeded the United States to firmly secure the top spot globally. During the same period, the weekly invocation volume of US large AI models was 4.28 trillion Tokens. The top six models by global invocation volume were all Chinese large AI models: DeepSeek-V4-Flash ranked first for the seventh consecutive week with 5.34 trillion Tokens; Xiaomi MiMo-V2.5 ranked second with 4.38 trillion Tokens; MiniMax M3 ranked third with 4.11 trillion Tokens; Tencent Hy3 preview ranked fourth with 3.13 trillion Tokens; Zhipu GLM-5.2 broke into the top five with 2.58 trillion Tokens; and StepFun Step 3.7 Flash ranked sixth with 1.55 trillion Tokens.

IV. Introduction to LLM Concept Stocks

NewTimeSpace Disclaimer: All content herein is the original work of NewTimeSpace. Any reproduction, reprinting, or use of this content in any other manner must clearly indicate the source as "NewTimeSpace". NewTimeSpace and its authorized third-party information providers strive to ensure the accuracy and reliability of the data, but do not guarantee the absolute correctness thereof. This content is for reference only and does not constitute any investment advice. All transaction risks shall be borne by the user.