Tencent pivots to pragmatic AI strategy with Hy3 model

🔑 Enhanced Key Takeaways

•Tencent's increased AI infrastructure spending, projected to significantly boost in the second half of 2026, is partly attributed to an improving domestic chip supply in China, potentially easing previous hardware bottlenecks.
•The Hy3 preview model is a Mixture-of-Experts (MoE) architecture that integrates 'fast and slow thinking' capabilities, allowing it to route routine queries to quick pattern-matching experts and complex problems to deeper reasoning chains.
•Tencent has open-sourced the Hy3 preview model on platforms such as Hugging Face, GitHub, and ModelScope, and offers its API access at a cost-effective price, approximately one-tenth of GPT-4-class rates.
•The recent AI organizational restructure involved the appointment of former OpenAI researcher Yao Shunyu as Chief AI Scientist, reporting directly to Tencent President Martin Lau, and the disbanding of the Tencent AI Lab, with its personnel integrated into the Hunyuan large-model team.
•The development of Hy3 preview was deeply integrated with Tencent's product teams, including Yuanbao, WorkBuddy, CodeBuddy, ima, and QQ Browser, with live product metrics directly influencing training priorities to ensure real-world applicability.

📊 Competitor Analysis▸ Show

Competitor Analysis: Tencent Hy3 vs. Key AI Models

Tencent's Hy3 preview model enters a competitive landscape, particularly within China, where it faces strong domestic players like Alibaba, Baidu, and ByteDance, as well as global leaders such as OpenAI and Google. Tencent's strategy emphasizes engineering efficiency and real-world application, contrasting with some competitors' focus on raw parameter counts.

Feature/Metric	Tencent Hy3 Preview	Alibaba Qwen2.5-Max	Baidu Ernie Bot 4.0	Cohere Command A	OpenAI GPT-4-class (Reference)
Model Type	Mixture-of-Experts (MoE)	-	-	Mixture-of-Experts (MoE)	-
Total Parameters	295B (21B active)	-	-	111B	Trillions (estimated)
Context Window	256K tokens	-	-	256K tokens	-
Input Price (per 1M tokens)	~$0.066	-	-	-	~$0.66 (estimated, 10x Hy3)
Output Price (per 1M tokens)	~$0.26	-	-	-	-
SWE-bench Verified	74.4%	-	-	-	~80.8% (Claude Opus 4.6)
Other Benchmarks	Top on OpenRouter for tool calling & programming; excels in STEM, context learning, instruction following, coding, agent tasks.	Outperforms DeepSeek V3 in Arena-Hard, LiveBench, LiveCodeBench, GPQA-Diamond.	Topped Chinese LLM lists; lagged GPT-4/Claude-3 in semantic comprehension, coding.	Excels in agentic, multilingual, and coding use cases.	Leading performance across various benchmarks.
Key Focus	Engineering efficiency, real-world application, agentic workflows, cost-efficiency.	Open-source, strong performance, image recognition.	Multi-modal capabilities, Chinese language focus.	Agentic, multilingual, coding, hardware efficiency.	General-purpose, cutting-edge capabilities.
Ecosystem Integration	WorkBuddy, CodeBuddy, Yuanbao, WeChat, Tencent Docs, QQ.	Cloud services, Qwen family open-source models.	Ernie Bot, integrated with Baidu ecosystem.	-	Various enterprise and consumer applications.

Note: Pricing and benchmark data for competitors are not consistently available or directly comparable across all search results. The GPT-4-class pricing is an estimation based on Tencent's claim of being one-tenth the cost.

🛠️ Technical Deep Dive

Model Architecture: Hy3 preview is a Mixture-of-Experts (MoE) model designed to integrate both 'fast and slow thinking' capabilities. It routes routine queries to quick pattern-matching experts and complex problems to deeper reasoning chains.
Parameter Count: The model has a total of 295 billion parameters, but only 21 billion active parameters are utilized per forward pass, contributing to its efficiency. It also features 3.8 billion MTP (Multi-Task Pre-training) layer parameters.
Context Window: Hy3 preview supports a substantial context window of up to 256,000 tokens.
Inference Efficiency: The model delivers a 40% improvement in inference efficiency, achieved through deep co-optimization between its architecture and inference framework, alongside enhancements across the inference stack, including compute performance and quantization algorithms.
Reasoning Capabilities: It supports configurable reasoning levels (disabled, low, and high modes), allowing it to balance speed and depth based on task requirements. It has demonstrated strong performance on challenging STEM benchmarks like FrontierScience-Olympiad and IMOAnswerBench, and achieved excellent results in the Tsinghua Qiuzhen College Math PhD qualifying exam (Spring '26) and the China High School Biology Olympiad (CHSBO 2025).
Agentic Workflows: Hy3 preview is specifically designed for agentic workflows and production use, capable of reliably powering complex agent workflows of up to 495 steps. It supports integration with popular open-source agent frameworks such as OpenClaw, OpenCode, and KiloCode.
Coding Performance: The model shows significant gains in coding, achieving competitive scores on mainstream coding agent benchmarks like SWE-bench Verified (74.4%), Terminal-Bench 2.0 (54.4%), and search agent benchmarks like BrowseComp (67.1%).
Training Infrastructure: Hy3 preview is the first model trained on Tencent's rebuilt AI infrastructure, which was re-engineered around principles of capability systematization, evaluation authenticity, and cost-efficiency.
Cost-Effectiveness: On Tencent Cloud TokenHub, Hy3 preview offers competitive pricing, with input costs starting at approximately USD 0.18 per million tokens (or RMB 1.2 per million input tokens) and output costs starting at approximately USD 0.59 per million tokens.
Research on CALM: Tencent is also exploring Continuous Autoregressive Language Models (CALM), a novel approach that predicts the next vector of meaning instead of individual tokens, aiming to dramatically reduce sequence length, attention costs, and overall compute requirements.

🔮 Future ImplicationsAI analysis grounded in cited sources

Tencent's 'engineering-first' AI strategy and focus on real-world application performance with models like Hy3 will likely accelerate the integration of AI into its vast product ecosystem.

The Hy3 model was developed in a single loop with product teams like Yuanbao, WorkBuddy, and CodeBuddy, with live product metrics shaping training priorities, indicating a strong product-market fit focus and rapid deployment across Tencent's services.

Tencent's significant increase in AI infrastructure investment, driven by improving domestic chip supply, positions it to strengthen its competitive stance against both domestic and international AI rivals.

Tencent plans to substantially boost AI infrastructure spending in the second half of 2026 as Chinese chipmakers ramp up domestic AI chip production, which could alleviate previous supply constraints and enable faster development and deployment of advanced AI models.

The open-sourcing of Hy3 preview and its competitive API pricing will likely foster broader adoption and development of AI agents and applications within the Chinese developer ecosystem.

Tencent open-sourced Hy3 preview on platforms like Hugging Face and GitHub and offers API access at roughly one-tenth of GPT-4-class rates, making it highly accessible and attractive for developers and enterprises to build new AI-powered solutions.

⏳ Timeline

2016

Tencent AI Lab established, focusing on fundamental AI research.

2023-09

Tencent debuts its universal large language model, Hunyuan (Hy1), with over 100 billion parameters.

2025-12

Tencent undergoes a major AI organizational restructure, appointing Yao Shunyu as Chief AI Scientist and establishing dedicated AI Infra, Data, and Data Computing Platform departments.

2026-02

Tencent Cloud launches a closed beta of WorkBuddy, a desktop AI agent designed for autonomous workplace tasks.

2026-03

Tencent folds its prominent AI Lab into the Hunyuan large-model team, consolidating AI efforts.

2026-04

Tencent launches and open-sources the Hy3 preview model, emphasizing engineering efficiency and real-world application.

Tencent pivots to pragmatic AI strategy with Hy3 model

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

Competitor Analysis: Tencent Hy3 vs. Key AI Models

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (17)

👉Related Updates

Amazon Quick Launches Agentic Catalog Experience for Data Curators

Autonomous Intelligence and the End of Productivity Software

AI models rely on human-curated Wikipedia data

Rethinking AI Reliability: Moving Beyond Zero-Error Myths