Tencent has pushed its in-house AI family a big step up the ladder. On November 4, 2026, the company announced Hy4, the latest generation of its Hy large language model, and said it now sits among the strongest open-source models in the world. According to Tencent’s announcement, the model weights and system cards will follow “in the coming days,” which would put a model with vision, computer use and long-horizon agent skills into the hands of developers who want to run it themselves.
For businesses watching the open-model race, Hy4 is notable less for any single benchmark than for how tightly it has been shaped around real products: office work, financial research, game development and scientific workflows. Here is what Tencent says is new, and why it matters.
What is new in Hy4
According to Tencent’s press release, Hy4 is a substantial upgrade on the earlier Hy4 Preview. The company says it scaled up both pre-training and post-training, and the headline additions are:
- Vision for the first time. Hy4 can now understand images, which Tencent says enables fine-grained image processing, complex chart analysis and mathematical visual reasoning.
- Computer use. The model can operate both command-line tools and graphical interfaces. Tencent says it received specialized training on professional software such as Blender and CAD tools to help it plan and carry out long, multi-step tasks.
- Better long-context and agent skills. The release highlights gains in instruction following, long-document understanding and long-horizon tasks such as deep research.
Tencent adds that, despite these gains, Hy4 keeps the same API price as the preview, a point we return to below.
Built around real products, not just leaderboards
A recurring theme in the announcement is what Tencent calls deep model-product co-design: the model was tuned alongside the apps that use it.
Office work. On Tencent’s WorkBuddy assistant, the company reports an average evaluation score of 83.18 and a pass rate of 96.15 percent, with Hy4 ranking first in nine of twelve scenarios in the mainland Chinese market. Connected to WeCom, Hy4 can draw on chats, emails, documents, calendars and meetings to track project history, flag meetings that need preparation and surface approaching deadlines. A team “Projects” feature for WorkBuddy is on the way, Tencent says.
Financial analysis. Tencent says it trained Hy4 on thousands of specialized workspaces and tens of thousands of tasks drawn from fund research, securities analysis and quantitative strategy. In a blind evaluation run with Tencent Financial Technology, 213 financial professionals rated performance across 579 real tasks. Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash-1017 (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68), according to the company. Those are Tencent’s own numbers, so independent testing will be worth watching once the weights are public.
Fewer made-up answers. Tencent says targeted search optimization cut Hy4’s hallucination rate in internal tests from 12.95 percent in the preview to 7.2 percent, while errors caused by unsupported assumptions fell from 23.1 percent to 5.6 percent.
Front-end and game development. Working with developers in its Interactive Entertainment Group, Tencent integrated Hy4 with Unreal Engine, Unity and Godot. The company reports improvements of 19.1 percent on a GameDev benchmark and 9.8 percent on GameLogic compared with the preview.
A model that helps build models
One of the more forward-looking claims concerns research. Tencent says Hy4 was validated on TenureTrail, a set of 123 executable tasks across seven disciplines, from mathematical construction to physical modeling and socio-economic decision-making.
The company also says Hy4 is moving into what it calls recursive self-improvement. Using a harness named Hyra 2.0, Hy4 helped optimize the pre-training infrastructure for smaller models that share its architecture. Tencent has also open-sourced two foundation models developed with Hy4 and Hyra 2.0: HyCell, for single-cell representations, and HyPDE, for physics fields. These are aimed at giving researchers reusable building blocks rather than finished applications.
Pricing, availability and the open-weight plan
According to the announcement, Hy4 is available now through Tencent Cloud TokenHub and Tencent products including WorkBuddy, CodeBuddy, Yuanbao, WeCom’s AI assistant and ima, with API access also offered through third-party platforms such as OpenRouter. Integration into QQ, Tencent News, WeRead and some of Tencent’s games is under way.
Despite the upgrade, Tencent kept the preview’s pricing: roughly US$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens. WorkBuddy and CodeBuddy are free from November 4 to 15, and Hy4 is half price on TokenHub through December 31, 2026. A cheaper, faster Hy4 Flash is “scheduled for release soon.”
On the open-source side, Tencent says weights and system cards are coming within days and that the release process includes internal safety reviews focused on cybersecurity and biosafety. Its own biosafety review describes Hy4’s open-ended experimental design abilities as more limited than frontier models.
What it means for businesses and developers
- More choice at the top of open AI. If the weights land as promised, teams will have another capable, multimodal model they can host on their own infrastructure, which matters for data control and cost.
- Agents that use software, not just chat. Computer use plus training on tools like Blender and CAD points to assistants that can carry out tasks inside real applications.
- Price pressure continues. Keeping the same price after a major upgrade signals that competition among Chinese and global model providers is still pushing costs down.
- Test before you trust. Most performance figures are Tencent’s own. Run your own evaluations on the tasks you care about, especially for finance or research work.
The bottom line
Hy4 shows Tencent treating its model as the engine of a whole product ecosystem, from office assistants to game engines, while also preparing an open release. The real test comes when the weights are public and independent developers can measure it for themselves. Until then, Hy4 is a clear sign that the open-model field is getting more crowded, more capable and more affordable.