Tencent expands global availability of hybrid-MoE model Hy3 across products and cloud

Tencent expands global availability of hybrid-MoE model Hy3 across products and cloud

Tencent has announced the global expansion of its AI model Hy3 (formerly Tencent Hunyuan), which was officially launched on July 6. The model is now accessible through platforms such as WorkBuddy, Miora, and TokenHub, with a free trial for global users via WorkBuddy until August 31 (Pacific Time). Developers can also access Hy3 through APIs or open-source communities like Hugging Face and ModelScope. Hy3 is built on a hybrid mixture-of-experts (MoE) architecture that combines fast and deep reasoning, featuring 295 billion total parameters, 21 billion active parameters, and a maximum context length of 256K tokens. According to Tencent, Hy3 delivers performance comparable to models with two to five times more parameters in inference, code generation, and agent capabilities.

This expansion marks a strategic push by Tencent to position Hy3 as a global player in the foundation model market, leveraging both open-source distribution and enterprise cloud channels. Within a week of launch, Hy3 recorded over 68 times more API calls than its predecessor and ranked first on OpenRouter's global LLM usage chart. It is already integrated into third-party platforms such as Hermes, Kilo, Klein, and Cherry Studio. The model is available under the permissive Apache 2.0 license, with API pricing starting at $0.1288 per million input tokens and $0.5336 per million output tokens on OpenRouter. Tencent is also partnering with regional players like Cafe24 in Korea and Metelix in Japan to offer Hy3 through local channels.

For the AI industry, this move reinforces the hyperscaler distribution moat that Chinese tech giants are building around open-weight models. By combining a competitive open-source model with aggressive pricing, global cloud reach (66 availability zones across 23 regions), and integrations into productivity tools like WorkBuddy and creative tools like Miora, Tencent is challenging Western incumbents in the open-model race. The hybrid MoE design and the performance-to-parameter ratio highlight the efficiency gains that are becoming a recurring pattern in the foundation model segment. Tencent's claim of a 90% success rate and 34% faster task completion in WorkBuddy underscores the practical enterprise value, signaling a shift from pilot experimentation to production deployment.

As an analyst note, Hy3's rapid adoption on OpenRouter and its Apache 2.0 licensing position it as a credible alternative to models like DeepSeek and Qwen in the open-weight ecosystem. The global expansion through cloud partnerships not only increases its reach but also reinforces the trend of Chinese AI models becoming integral to international enterprise workflows. Tencent's ability to leverage its cloud infrastructure and product ecosystem gives it a unique advantage in scaling AI adoption across diverse markets, while the low-cost API pricing could pressure competitors on price-performance benchmarks. This is a strategic play that combines open-source community engagement with enterprise monetization, a hybrid approach that is likely to become a template for other large Chinese technology firms.

#Tencent #Hy3 #FoundationModels #HybridMoE #OpenSourceAI #CloudAI