๐ญ๐ฐSCMP TechnologyโขStalecollected in 0m
Ex-Tencent AI Lead: China Trailing in LLM Race
๐กUnderstand the leadership shifts and strategic hurdles facing China's top AI labs in the global LLM race.
โก 30-Second TL;DR
What Changed
Liu Wei, a key figure in Tencent's AI strategy, has left the company.
Why It Matters
This departure signals potential internal shifts or strategic re-evaluations within Tencent's AI division. It underscores the difficulty of maintaining top-tier AI talent in a highly competitive global landscape.
What To Do Next
Monitor Tencent's upcoming AI research papers and model releases to see if the leadership change impacts their technical trajectory.
Who should care:Founders & Product Leaders
Key Points
- โขLiu Wei, a key figure in Tencent's AI strategy, has left the company.
- โขThe Hunyuan model was launched by Tencent to compete in the generative AI space.
- โขIndustry experts suggest China faces structural hurdles in the global LLM race.
๐ง Deep Insight
Web-grounded analysis with 23 cited sources.
๐ Enhanced Key Takeaways
- โขLiu Wei, the former head of Tencent's Hunyuan team, departed in November 2024 to establish his own AI video generation startup, Video Rebirth, which has successfully secured $80 million in funding.
- โขTencent's Hunyuan model has been extensively integrated into over 700 of the company's internal business applications and scenarios, providing diverse capabilities including text-to-image and video generation.
- โขIn a strategic move to accelerate its generative AI push, Tencent launched its next-generation Hunyuan 3.0 large language model in April 2026 and is actively developing an advanced AI agent for its super-app WeChat.
- โขThe Chinese LLM market is embroiled in a fierce price war, with Tencent significantly cutting prices for its Hunyuan LLM in May 2024, making the lite version free and reducing standard versions by 50% to 87.5%.
- โขChinese open-source LLMs, including those from Tencent, have dramatically increased their global market share, growing from 1.2% in late 2024 to an estimated 30% by the end of 2025, surpassing the U.S. in global AI model downloads for the first time.
๐ Competitor Analysisโธ Show
| Company/Model | Key Features/Focus | Pricing Strategy (as of May 2024) | Benchmarks/Performance Notes |
|---|---|---|---|
| Tencent Hunyuan | Multimodal (text-to-text, image, video), in-context learning, agent usability, efficient MoE architecture (Hunyuan-A13B) | Lite version free; standard versions 50-87.5% price cut; Pro version 70% input price cut. | Hunyuan-A13B rivals OpenAI's o1 and DeepSeek's R1, 2.2-2.5x faster inference than larger models like Qwen3-A22B. Hunyuan-A13B-Instruct outclasses Qwen3-A22B in agentic reasoning. |
| Baidu Ernie Bot | General-purpose LLM | Lite version free. | Competes in the general chatbot market. |
| Alibaba Qwen | Multilingual, multimodal (text, image, audio, video), progressive generation (Qwen-VLo), MoE architecture (Qwen 2.5-Max) | Participated in price war. | Qwen 2.5-Max shows impressive performance in math, programming, and general tasks, competing with DeepSeek-V3 and GPT-4o. Qwen family surpassed 700 million downloads by Jan 2026. |
| ByteDance Doubao | China's most popular AI chatbot, deeply integrated with TikTok/Douyin, personalization, multimodal | Aggressive pricing (99.8% lower than OpenAI's GPT-4 for enterprises). | Over 100 million daily active users by end of 2025. |
| DeepSeek | Coding powerhouse (DeepSeek V3), cost-efficient, multimodal (DeepSeek V4) | Aggressive pricing. | DeepSeek V3 achieves strong scores across benchmarks, excels in code generation and debugging. DeepSeek-R1 achieved high performance with fewer resources and lower production cost. |
| iFlyTek Spark LLM | General-purpose LLM, strong in voice AI | Significantly reduced prices for some versions; lite version free. | Leading in voice AI race. |
๐ ๏ธ Technical Deep Dive
- Hunyuan-A13B: Utilizes a Mixture-of-Experts (MoE) architecture with a total of 80 billion parameters, but only 13 billion are actively used during inference, optimizing for efficiency.
- Hunyuan-T1: Described as the world's first ultra-large-scale Hybrid-Transformer-Mamba Mixture-of-Experts (MoE) model, combining Transformer and Mamba architectures for faster training and lower inference costs. It achieves 2X faster decoding than traditional Transformer models.
- Hunyuan-T1 Features: Includes dual-mode reasoning, allowing users to switch between 'slow-thinking' (chain-of-thought) and 'fast-thinking' modes, and employs a unified reward system combining self-rewarding and human preference alignment (RLHF).
- Hunyuan Image 3.0: An open-sourced text-to-image model with 80 billion total parameters and 13 billion active parameters during inference. It integrates an MoE architecture with the Transfusion method to unify multimodal understanding and generation capabilities.
- Hunyuan Image 3.0 Training: Employs a progressive training strategy that includes pre-training (low to high resolution, low to high quality), instruction tuning with chain-of-thought data, supervised fine-tuning with high-quality data, and reinforcement learning using DPO and GRPO algorithms.
- Hunyuan 3D Engine (Hunyuan3D-2.1): Features a two-stage generation pipeline for 3D assets, first creating a bare mesh and then synthesizing a texture map. It is fully open-sourced and incorporates Physically-Based Rendering (PBR) texture synthesis for photorealistic material generation.
- Inference Optimization: Tencent Hunyuan collaborated with NVIDIA to develop the AngelHCF inference engine, built on TensorRT-LLM, which has optimized inference costs by over 90% and reduced text-to-video inference time to under 2 minutes for Hunyuan models.
- Training Data Scale: Hunyuan-A13B was pretrained on an extensive dataset of 20 trillion tokens.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
Tencent will intensify its focus on integrating AI agents across its vast product ecosystem.
The development of an advanced AI agent for WeChat and the emphasis on agent usability in Hunyuan models suggest a strategy to embed AI deeply into its existing user base and enterprise tools.
The fierce price war in China's LLM market will continue, driving consolidation and pushing smaller players towards niche applications or acquisition.
Major tech giants like Tencent, Baidu, and ByteDance have aggressively cut prices, making it challenging for startups to compete on cost and scale, forcing them to find specialized value propositions.
China's open-source LLM ecosystem will continue to gain global market share, particularly in cost-efficient and specialized models.
Chinese open-source models have already surpassed U.S. downloads, and their focus on efficiency (e.g., MoE, lower active parameters) and aggressive open-sourcing strategy appeals to developers and enterprises seeking lower inference costs.
โณ Timeline
2012
Liu Wei received his Ph.D. in Computer Science and Electrical Engineering from Columbia University.
2016
Liu Wei joined Tencent as a Distinguished Scientist.
2017
Liu Wei became Director of the AI Lab Computer Vision Center and Director of Ads Multimedia AI at Tencent.
2024-11
Liu Wei departed Tencent and founded his AI video generation startup, Video Rebirth.
2025-11
Tencent globally launched its Hunyuan 3D creation engine.
2026-04
Tencent launched its next-generation Hunyuan 3.0 LLM and updated its Hunyuan AI model (Hy3).
๐ Sources (23)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
๐ฐ
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: SCMP Technology โ