Tencent Releases Hy4 Model With Computer Use, Keeping Preview Pricing

The launch follows Hy4 Preview, which came out on August 28. The preview has 770 billion parameters, of which 49 billion are active at a time. Tencent published the preview's weights on Hugging Face under an Apache 2.0 licence, which allows commercial use. Both versions have a context window of 1 million tokens. SiliconANGLE said that puts Tencent alongside the open-weights frontier models from Mistral AI, Moonshot and Z.AI.

Tencent said it plans to release Hy4's weights and system cards in the coming days, and a cheaper Flash version is due soon.

What Has Changed Since the Preview

Tencent said it scaled up both pre-training and post-training for the final release. It reports gains over the preview in instruction following, long-context understanding, long agentic tasks and deep research. The company also said it found the previous version spent too much time reasoning and re-checking its own work in its "thinking" mode.

The model is aimed at everyday office and professional work. Tencent highlighted web development, document summarisation, visual creative tasks and presentation building. It also said Hy4 is better at front-end and game development, including complex business logic, state management and orchestration.

For games, Tencent said Hy4 integrates closely with the Unreal Engine, Unity and Godot engines and can support complex game logic and multiplayer features. The company reported improvements of 19.1 per cent on the GameDev benchmark and 9.8 per cent on GameLogic compared with the preview. It has not released the underlying scores, so the gains cannot be checked independently.

Tencent also described work on what it calls recursive self-improvement. It said an autonomous AI research agent system called the Hyra 2.0 harness used the model to optimise the pre-training infrastructure for smaller models built on the same architecture.

Availability and Pricing

Hy4 is available through Tencent's own products and platforms, including Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, the Yuanbao assistant, WeCom's AI Assistant and Ima. Developers can also reach it through third-party services such as OpenRouter, Hermes and OpenCode.

Tencent kept the preview's prices: about $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens.

Why It Matters

Hy4 is another sign that Chinese labs are competing on agent features as well as on price. Computer use, where a model reads a screen and takes actions in applications, has so far been led by US labs. A Chinese model offering it at under $3 per million output tokens, and with open weights promised, adds to the price pressure that open models from China and Europe already put on Western providers.

The gaming focus is notable. SiliconANGLE pointed out that the release comes shortly after Anthropic blocked Claude Code from helping users clone existing video games. That leaves room for rivals to market themselves to game developers, although Tencent has not said what limits, if any, Hy4 places on that kind of use.

For enterprise buyers, including in the Gulf, where demand for open-weights models that can be hosted locally is growing, the system cards and weights will be the more important test. Until they are out, the company's own descriptions are the only basis for comparing Hy4 with other models.