Beating AI says Tencent revealed on Reddit that the next-generation Hy model will significantly expand parameters, and Hy4-mini and Flash-Lite are also in the plan.
Dongcha Beating AI stated that Tencent Hunyuan has open-sourced three versions of Hy4 flash, and added that the flagship Hy4 with 770B parameters is still being accelerated for open source, with no specific timeline announced.
Tencent released Hy4 flash, with input, output, and cache-hit prices reduced by 12 times, 9 times, and 15 times respectively compared with the Hy4 list price. Official tests show that its cost for PPT collaborative editing tasks is about one 16th of Hy4's, but its scores for terminal operations and programming are noticeably lower.
The joint project team of Tencent Hunyuan and WorkBuddy stated that after the official version of Hy4 was integrated, quota limits and task queuing occurred on the first day, so the inference cluster has been scaled up by a factor of 10. Due to limitations in high-end computing power and peak requests, quotas may still apply or queuing may occur during certain periods.
Pang Tianyu, Tencent Hunyuan's person in charge of multimodal reinforcement learning, reportedly has left the company and may start a business next; the relevant information has not yet been officially confirmed.
Tencent Hunyuan has launched Hy Image3.5 preview, increasing the reference-image limit to 20 and supporting multi-round editing and direct 4K output. The model is now available on Miora and is free for a limited time from September 22 to October 7.