NEWS → · Models · Tencent
Tencent launches Hy4 — a 770B flagship with vision and computer use; open weights promised in the coming days
Nov 4, 2026
Announced by Tencent on Nov 4, 2026: Hy4, the new flagship in its Hunyuan line, is live today across Tencent's own products and cloud — the first in the line that can see images and operate a computer. The open weights are not out yet: Tencent says they arrive, with a system card, in the coming days.
Image: Tencent via tencent.com
What launched, from Tencent's own release of Nov 4, 2026: Hy4 adds visual understanding to the Hunyuan line for the first time — fine-grained image processing, complex chart analysis and mathematical visual reasoning — and computer-use capabilities, driving both command-line and graphical interfaces, with specialised training on professional software such as Blender and CAD. It is available now on Tencent Cloud TokenHub, in WorkBuddy, CodeBuddy, Yuanbao, ima and WeCom's Dayuan assistant, and its API is being made available via OpenRouter, Hermes and OpenCode.
The scale, in Tencent's launch figures: 770 billion total parameters in a mixture-of-experts design, about 49 billion active per token — the full version of the model previewed in August. Tencent ranks it among the world's leading open-source models on its own blind financial evaluation (213 professionals scoring real tasks: Hy4 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97 and Kimi K3.1 Pro at 2.68). That is Tencent's internal evaluation, reported as such — independent benchmarks will take time.
The catch, stated plainly: there is nothing to download today. Tencent says the model weights and system card will be released in the coming days, and no licence has been published yet — so despite the open-source framing, Hy4 cannot be self-hosted, fine-tuned or inspected until the files actually appear. A cheaper Hy4 Flash is also scheduled, with no date given. We have added Hy4 to the model catalogue with its status marked exactly that way.
The price of using it hosted, from Tencent's release: about $0.90 per million input tokens, $2.70 per million output and $0.045 per million cached input — the same pricing as the Preview — with TokenHub at 50% off through Dec 31, 2026 and WorkBuddy/CodeBuddy free from Nov 4 to 15. For GPU renters the number that matters comes later: at 770B parameters, the plain arithmetic says even a 4-bit build of the weights is about 385GB, roughly 462GB with working headroom — at least five 96GB cards or six 80GB cards when the weights land. Until then, Tencent's API is the only way in.
What it means if you rent GPUs
Another frontier-scale model joins the open-weight queue — and, as with Kimi K3.1 Pro, the honest headline is the gap between the announcement and the download. When the Hy4 weights do appear, the deployment maths is cluster maths, not single-card maths, and the licence that ships with them will decide who can actually use them commercially. Both are worth checking here before planning around them.