Weekly Report on Hong Kong-Listed LLM Concept Stocks (June 22 - 26, 2026)

I. Market Overview

This week (June 22-26), the Hong Kong stock market faced intense broad-market selling pressure. The Hang Seng Index (HSI) fell approximately 5.2% over the week, while the Hang Seng TECH Index tumbled more than 7.5%. Equities across the LLM and AI supply chains mirrored the broader market's high volatility, demonstrating an initial rally early in the week before succumbing to a sharp correction.

Early Week (June 22):AI concept stocks staged a powerful counter-trend breakout. MINIMAX-W skyrocketed 23.89%, KNOWLEDGE ATLAS soared 15.09%, and MININGLAMP-W surged 14.30%.

Mid-to-Late Week (June 23-26):Market sentiment reversed abruptly. On June 23, under pressure from impending lock-up expirations in July, MINIMAX-W plummeted 16.46% and KNOWLEDGE ATLAS dropped 9.96%. On June 24, although TSMC's advanced process price-hike projections fueled a counter-trend surge in SMIC and Hua Hong Semiconductor, the AI application segment diverged further, with MINIMAX-W shedding another 7.26%. On June 25, KNOWLEDGE ATLAS rebounded 8.10% due to the technical premium of its newly launched GLM-5.2 model, but large-cap internet tech equities overall dipped to over one-year lows. On June 26, a final wave of unilateral liquidations hit the market; KNOWLEDGE ATLAS fell 12.94% to record its steepest single-day drop of the week, while MINIMAX-W and SENSETIME-W shed 6.48% and 5.88% respectively.

Capital Flows:Southbound capital experienced highly volatile swings this week, printing a net outflow of HKD 5.82 billion on Monday, reversing to net inflows of HKD 10.371 billion and HKD 15.751 billion on Tuesday and Wednesday, before converting back to deep net outflows of HKD 8.874 billion and HKD 2.504 billion on Thursday and Friday.

II. Weekly Performance of LLM Concept Stocks

Data Source: HKEX. For reference only, not investment advice. Data as of market close on June 26, 2026.

Weekly Summary Notes:

MININGLAMP-W (+3.31%)emerged as the sole LLM concept stock to secure positive returns this week. Capital pools continuously validated its unique positioning as the premier "Agentic AI" pure-play and its vertical domain multi-model collaboration ("Scaling Out") delivery architecture.

KNOWLEDGE ATLAS (-2.29%)demonstrated robust defensive attributes, heavily insulated from deeper losses due to the ongoing technology dividends generated by its flagship GLM-5.2 release.

XUNCE (-23.13%)significantly underperformed the sector, succumbing to intense unilateral liquidation pressure.

BABA-W (-14.68%) and MINIMAX-W (-14.19%)rounded out the bottom tier. MINIMAX-W was persistently weighed down by psychological overhang regarding its impending July 9 lock-up expiration, which unlocks approximately 63% of the company's total outstanding share capital.

III. Macro and Industry News Review

1. Global LLM Call Volumes Rise for Ninth Week; Chinese Models Sweep Top Four Positions

According to OpenRouter data, global AI large model call volume hit 46.7 trillion tokens last week (June 15–21), up 4.7% week-on-week for its ninth straight weekly expansion. Weekly token volume for Chinese models reached 18.81 trillion tokens, outperforming the U.S. for the eighth consecutive week to solidify its global lead. Chinese architectures dominated the top four spots: DeepSeek-V4-Flash retained the number one position for the fifth consecutive week (4.94 trillion tokens), followed by Xiaomi's MiMo-V2.5, MiniMax M3, and Tencent's Hy3 preview. DeepSeek models aggregated 8.65 trillion tokens, capturing an 18.5% global market share.

2. DeepSeek Debuts on Hurun Unicorn Index with RMB 340 Billion Valuation, Launching Aggressive Hiring Drive

On June 25, DeepSeek ranked 12th on the 2026 Global Unicorn Index with an estimated valuation of RMB 340 billion, leading all newly listed unicorns globally. Concurrently, the firm launched a massive recruitment campaign aimed at doubling the headcount across all operational departments. Previously, on June 23, DeepSeek established its specialized "Harness" group, operating under the strategic engineering thesis of "Model + Harness = Agent" to build native developer tool ecosystems.

3. Five Chinese Ministries Jointly Issue Action Plan to Promote "AI + Transportation" Ecosystems

On June 25, ministries including the Ministry of Transport jointly issued the Action Plan for the Innovation of Typical Application Scenarios of "Artificial Intelligence + Transportation". The document explicitly mandates the R&D and testing of "end-to-end" large models for autonomous driving across freight and logistics corridors, encourages civil aviation vertical LLMs, and supports AI-driven multi-modal travel planning systems.

4. OpenAI Launches First Custom In-House AI Inference Chip "Jalapeño"

On June 24, local time, OpenAI partnered with Broadcom to introduce "Jalapeño," its first custom AI inference ASIC designed explicitly to optimize LLM execution pipelines. Engineering prototypes are currently undergoing internal laboratory benchmark assessments, with commercial deployment into gigawatt-scale data center clusters scheduled for late 2026.

5. CAC: China Automatically Deploys More Than 900 Large Models for Public and Enterprise Use

At the opening ceremony of MWC Shanghai 2026 on June 24, the Cyberspace Administration of China (CAC) revealed that over 900 large models have been formally deployed across China, spanning manufacturing, agriculture, and life sciences. Regulators are actively constructing compliance standards for AI agent applications and synthetic content identification.

6. Volcano Engine: Daily Volume of Doubao LLM Blows Past 180 Trillion Tokens

On June 23, Volcano Engine disclosed that its Doubao LLM's average daily token usage surpassed 180 trillion tokens—a 1,500-fold expansion over two years—securing a 49.5% market share in China's public cloud MaaS segment. The group released Doubao 2.1 Pro, lowering token processing costs by nearly 80% relative to Claude Opus 4.6.

IV. Key Corporate Dynamics

MININGLAMP-W (02718.HK):Formally executed infrastructure integrations with major distribution platforms including WeChat and Tongyi Qianwen (Qwen). The company launched its enterprise-level "Agent Ready" solution across four critical vectors: intent comprehension optimization, recommendation tuning, dynamic business rule alignment, and cross-platform communication protocols.

SENSETIME-W (00020.HK):Disclosed at its general meeting that its enterprise-grade Token Plan has printed a 7-fold sequential monthly volume surge since its May launch. Its desktop agent, "Raccoon," reached 20 million historical users with Weekly Active Users (WAU) consolidating at 2.1 million.

TENCENT (00700.HK):Completed multi-terminal ecosystem deployment on June 23 with the official App Store release of its OS-level AI assistant, Marvis iOS. Concurrently, its native WeChat AI assistant, "Xiaowei," entered gray-box testing, utilizing its internal WeLM foundational model augmented by DeepSeek APIs.

BIDU-SW (09888.HK):Launched the Baidu Qianfan Token Plan Enterprise Edition on June 25, taking the lead in deploying direct API optimization and native adaptation pipelines for KNOWLEDGE ATLAS's flagship GLM-5.2 model.

BABA-W (09988.HK):Unveiled Qwen-AgentWorld on June 24, introducing the industry's first Native Language World Model (LWM) optimized natively across CPT, SFT, and RL pipelines to support simulated multi-agent text and GUI interaction environments.

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.