Tencent open-sources HunYuan Hy4 with 770B parameters, 1M context

Tencent's HunYuan Hy4 preview, open-sourced Aug. 28, packs 770 billion total parameters with 49 billion active per token and a 1 million-token context window, more than doubling the scale of its Hy3 predecessor and quadrupling its context length. The Mixture-of-Experts model targets real-world productivity tasks — coding, office work, game development and scientific research — rather than chasing general benchmark scores alone.

"We scaled Hy4 preview on three fronts: model size, context length, and training data," the Tencent Hy Team said in the model's GitHub release. "Stronger pre-training and a substantially larger post-training run compound into another step change in capability — the largest generation-over-generation gain we've measured."

Hy4 preview uses 78 layers with 256 routed experts and one shared expert per layer, activating the top eight per token. A native MTP layer of 10 billion parameters handles speculative decoding. The architecture borrows Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse index reuse, and identity Hyper-Connections to expand inter-layer information flow. The model is released under the Apache License 2.0, with an FP8 quantized version available on Hugging Face, ModelScope, GitCode and CNB.

The jump from Hy3 is steep: 295 billion total parameters and 21 billion active rose to 770 billion and 49 billion, while context length expanded fourfold from 256,000 tokens. Tencent said it built training data with in-house software engineers, game developers, finance analysts and security experts, then co-designed the model with products including WorkBuddy and CodeBuddy. In a blind side-by-side evaluation, 163 internal experts rated Hy4 preview slightly ahead of Zhipu AI's GLM 5.3 (2.99 vs. 2.92 average) and Moonshot AI's Kimi K3 (2.99 vs. 2.94) across 203 engineering tasks.

Where Hy4 preview lands in the open-source race

The release positions Tencent against a crowded field of Chinese open-source models, including Alibaba's Qwen, Baidu's Ernie and DeepSeek. Tencent said Hy4 preview can build a Three.js 3D website from scratch, generate playable game prototypes in Unity from a single prompt, and process 72 financial documents in one run to flag duplicate reimbursements and budget overruns. It is already integrated into WorkBuddy, CodeBuddy, Yuanbao, ima, Tencent Cloud TokenHub and OpenRouter, with a two-week free trial for the first two.

A notable feature: Hy4 preview is participating in its own development. Tencent said the model helps optimize training methods, data strategies, evaluation systems and underlying operators — proposing solutions, running experiments and feeding the resulting code and logs back into the next development cycle. Since rebuilding its infrastructure, Hunyuan has iterated major versions roughly every two months, shipping previews before official releases.

What it means for investors

The open-source push strengthens Tencent's cloud and AI business against rivals that monetize models through API access and enterprise tools. Tencent previously signaled in its financial report that Hy4 would be larger than Hy3 and that real product feedback would drive training. The model's integration into paid tools like WorkBuddy and CodeBuddy gives Tencent a path to convert model capability into software revenue, though the company has not disclosed pricing or inference cost per token. Tencent did not disclose the test conditions for its side-by-side evaluation, which relied on its own employees rather than independent benchmarks.

This article is for informational purposes only and does not constitute investment advice.