AIBRIEF
Hy4 preview is a 770B parameter open weight language model, free for two weeks on WorkBuddy and CodeBuddy; 'top tier' is Tencent's own internal study, not an independent benchmark.
Tencent has released and open-sourced Hy4 preview, a 770B-parameter MoE language model that activates 49B parameters per token and serves a >1M token context window. Tencent is calling the model the "top tier" of open-source models.
That superlative rests on a Tencent-internal blind study: 163 in-house experts graded 203 engineering tasks, with Hy4 preview scoring 2.99 out of 4.00 versus GLM-5.3 at 2.92 and Kimi K3 at 2.94. Neither comparator is a frontier model, and no independent benchmark reproduction, including MMLU, HumanEval, AIME, or LiveCodeBench, appears in the primary materials.
Readers can try it now across several surfaces. WorkBuddy and CodeBuddy are free for two weeks; Yuanbao, ima, and Tencent Cloud TokenHub are also wired in, and the OpenRouter listing is live. A Hacker News observer reported "trillions of tokens generated in a couple days" on OpenRouter, more than GLM 5.3 drew in a week, alongside community-noted cache pricing of 5%, well below the 10–20% range Explainx reports elsewhere at roughly $0.83 per million input tokens. Tencent's prior Hy3 remains free through Sept. 30.
What remains unknown: independent benchmark scores, the license terms attached to the open weights, and whether the early OpenRouter demand holds.