Tencent open-sourced its new flagship on August 28: Hy4 preview, a 770B-parameter MoE with only 49B active per token and a context window past 1M tokens. Apache 2.0, weights on Hugging Face. China’s open-weights war just got its heaviest entrant since DeepSeek.
What it actually is
A frontier-scale open model built for long-horizon work — coding, office tasks, research. The architecture is the interesting part: Gated DeepSeek Sparse Attention plus IndexCache, which reuses sparse indices across layers to keep 1M-token inference cheap. 256 routed experts, top-8 plus one shared. On DeepSWE it jumped from 28.0 to 64.3, and it beat DeepSeek V4 Pro on Terminal Bench 2.1 with 85.4. Tencent says the model optimized its own training stack, lifting end-to-end throughput 31.8%.
How to use the API
API access runs through Tencent Cloud TokenHub and OpenRouter at $0.83 per million input tokens, $2.50 output. That pricing plus 1M context makes it a real option for agentic coding and whole-repo analysis. Consumers get it inside WorkBuddy, CodeBuddy, and Yuanbao — free for two weeks. Or self-host: the weights are yours.
You Might Also Like
- Zenmux Just hit 1 on Product Hunt Heres why Everyones Paying Attention
- Tencent Hunyuan hy3 Preview Goes Open Source 295b moe 21b Active 256k Context
- Starnus Just hit 1 on Product Hunt and Yeah its Worth the Hype
- Lovon Just Topped Product Hunt on Valentines day and its not a Dating app
- Google Pomelli Photoshoot Just Dropped and Product Photographers Should be Paying Attention
Top AI Product 
Leave a comment