Tencent’s Hy3 model drives massive adoption and international integrations, featuring a 295B MoE architecture and hybrid reasoning capabilities.
Following its official open-source release on 6 July 2026, Tencent announced expanded global access for Hy3 across international products, workflows, and cloud services. The model is released under the commercially friendly Apache 2.0 Licence on open-source platforms including Hugging Face and ModelScope, and it is accessible globally through WorkBuddy (free until 31 August 2026), Tencent Design Miora, Tencent Cloud TokenHub, and API aggregators such as OpenRouter.
Hy3 is built on a Mixture-of-Experts (MoE) architecture incorporating a hybrid fast-and-slow-thinking reasoning framework. It features 295 billion total parameters, with 21 billion active parameters per token, and supports a context length of up to 256K tokens. Additionally, it offers a configurable reasoning_effort parameter (no_think, low, high), enabling a single model to handle both low-latency direct responses and deep chain-of-thought reasoning tasks.
API calls for Hy3 surged over 68 times compared to the previous-generation Hy2 model within its first week of launch, driving the model to the top of OpenRouter’s global LLM token usage leaderboard following its initial preview release. In practical applications, Hy3 delivered a task success rate exceeding 90 per cent within Tencent’s WorkBuddy app while reducing average task completion time by 34 per cent compared to previous model iterations. The model has already been adopted by third-party developer tools including Hermes, Kilo, Cline, OpenClaw, OpenCode, and Cherry Studio.
Pricing on OpenRouter starts at USD$0.1288 per million input tokens and USD$0.5336 per million output tokens. To accelerate global enterprise adoption, Tencent has collaborated with regional international partners, including South Korea’s e-commerce platform provider Cafe24 and Japan’s enterprise AI platform company Metelix, to integrate Hy3 into their respective AI gateway services.
