Tencent's Hy3 draws 50%+ queue rate, forcing emergency capacity expansion

Tencent Holdings Ltd.'s Hunyuan Hy3 large model triggered a computing capacity crunch on its WorkBuddy platform within hours of launch, with queue rates exceeding 50% as demand overwhelmed available resources.

"The usage surge exceeded our initial capacity projections," Tencent's Hunyuan team said in a statement, adding that additional computing resources were allocated and capacity expansion completed within hours.

Computing resource consumption peaked starting at 10 a.m. on July 8, with the queue rate surpassing 50% by that afternoon. Tencent said it promptly allocated available resources to stabilize services and has now completed the expansion.

The demand validates Tencent's strategy of positioning Hy3 as a cost-efficient alternative to larger models. At 1 yuan per million input tokens — roughly $0.14 — Hy3 undercuts many rival offerings, potentially accelerating enterprise adoption of Tencent's AI tools and driving incremental cloud revenue.

A Smaller Model Competing With Larger Flagships

Hy3 has 295 billion total parameters with 21 billion activated and a 256,000-token context window. That puts it well below flagship rivals such as GLM-5.1, which has 744 billion total parameters and 40 billion activated. Yet Tencent's published benchmarks place Hy3 in a comparable performance band on agent and reasoning tasks, challenging the assumption that bigger models are necessarily better.

The model's Apache 2.0 license allows commercial use without restrictions, and its API pricing — 1 yuan per million input tokens, 4 yuan per million output tokens, and 0.25 yuan per million cache-hit input tokens — positions it as a low-cost option for high-volume applications such as coding assistants, office agents and multi-step automation tools. Many competing models charge multiples of that rate for comparable capability.

Demand Tests Infrastructure as Adoption Accelerates

The surge suggests Hy3 is gaining traction in real-world deployment rather than benchmark leaderboards alone. DeepSeek's V4 Pro and Zhipu AI's GLM-5 series remain formidable competitors in China's open-model race, but Hy3's combination of permissive licensing, aggressive pricing and agent-optimized design gives Tencent a differentiated pitch. The question is whether Tencent can sustain the infrastructure to support demand as adoption scales.

Tencent released Hy3 Preview in April and moved to the official version roughly two months later, following a broader reconstruction of Hunyuan's model infrastructure. That pace reflects a faster development loop connecting model research with product testing and real-world feedback — a cycle that matters more as AI models move from chat interfaces into coding assistants, office agents and enterprise automation systems.

The capacity crunch also highlights the infrastructure demands of agent-optimized models. Agent tasks typically require more computing resources than standard chatbot interactions because they involve long-context understanding, tool calling, multi-step reasoning and error recovery. A model that performs well in isolated language tasks may still struggle when asked to operate across software environments, making real-world testing essential.

Tencent plans to make Hy3 available through its Tencent Cloud TokenHub and open-source communities including Hugging Face and ModelScope, as well as global developer platforms such as OpenRouter, Cline and CherryStudio. That broad distribution strategy matters because open-model adoption depends not just on benchmark performance but also on licensing, pricing, tool compatibility and ease of integration into existing workflows.

Tencent shares rose 0.5% on July 8, with short-selling volume at HK$4.09 billion representing 15.3% of turnover. CMSI maintained its overweight rating on Tencent, noting Hy3's entry into the top tier of AI models could accelerate the company's AI product lineup.

This article is for informational purposes only and does not constitute investment advice.