Tencent open-sources Hy4: 770B MoE, 1M context, $0.83/M input
The standout of the recent open-weight wave: frontier coding, a two-week free trial, and a model that tuned its own training for 31.8% more throughput.
Copy markdown
770B open weights, 49B active per token
Tencent open-sourced Hy4-preview: a 770B-parameter mixture-of-experts that fires just 49B params per token, with a 1M-token context window aimed at coding, office work, and research. Frontier scale you can now self-host or fine-tune.
Cheap on the API, free for two weeks
API access runs $0.834/M input, $2.501/M output and $0.042/M cached via Tencent Cloud TokenHub and OpenRouter — plus free use through WorkBuddy and CodeBuddy for a two-week window. Enough to wire it into an agent before committing.
It edges GLM-5.3 and Kimi K3 on engineering
In Tencent's internal eval (163 experts, 203 engineering tasks), Hy4 scored 2.99/4 against GLM-5.3's 2.92 and Kimi K3's 2.94. Vendor numbers on the vendor's tasks — worth a run on your own repo before you trust the ranking.
The model helped optimize its own training
Tencent says Hy4 tuned parts of its own training and inference stack, lifting end-to-end throughput 31.8% over baseline. That's infrastructure optimization, not self-improvement — but a hint of where the cost curve is heading.