Tencent Hunyuan releases Hy3 preview: 295B total / 21B active, Yao Shunyu's first model after returning to China

On September 14, 2026, Tencent Hunyuan released and open-sourced the Hy3 preview language model, the first model trained on Hunyuan's reworked infrastructure and the first model deliverable from chief AI scientist Yao Shunyu after his return to China. Hy3 preview is a fast-and-slow-thinking mixture-of-experts model with 295 billion total parameters and 21 billion active parameters per inference, supporting up to 256K context. Tencent said Hy3 preview delivers significant gains over its predecessor in complex reasoning, instruction following, in-context learning, coding, and agent capability, with strong placements across major international benchmarks. Hy3 preview has shipped first across Tencent Cloud, Yuanbao, ima, CodeBuddy, WorkBuddy, QQ, and Tencent Docs, and is available on the TokenHub model-as-a-service platform alongside support for OpenClaw, OpenCode, KiloCode, and other open-source agent products. Yao Shunyu said the team is open-sourcing the model to gather community feedback and iterate toward a production-ready release, and stressed that Hy3 preview is still early-stage, advising developers to evaluate carefully and keep fallback plans in production deployments. Tencent refreshed the TokenHub model matrix on the same day to cover Hy4 preview, Hy3, the DeepSeek V4 family, the Zhipu GLM family, the Kimi family, and the MiniMax family, exposing all of them via OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages protocols. The Hy3 preview release dovetails with the same day's OpenRouter data showing Tencent Hunyuan Hy4 preview at 16.8 trillion tokens of weekly traffic, the second-highest globally, as Tencent doubles down on a multi-model portfolio to compete for developer mindshare.