Tencent released a preview of its Hy4 flagship model on Friday, August 28, 2026, scaling its open-source offering to 770 billion total parameters with 49 billion activated per request. The model, posted to Hugging Face under the repository name tencent/Hy4-preview, carries an Apache 2.0 license and a reported 1 million token context window, positioning it alongside the largest open-weight systems from DeepSeek and Zhipu AI.
The release was first reported by Reuters and marks the third major iteration in Tencent's Hunyuan open-source lineage, following the April 2026 preview, the July 6 full Hy3 launch at 295 billion parameters, and the original Hunyuan-Large from October 2024 at 389 billion. Each iteration has roughly doubled the parameter count, though Hy4's architecture marks a significant departure from its predecessors, as analyzed in detail by explainx.ai.
Architecture: Dense First Layer, MoE Everything Else
Hy4 preview uses a Mixture-of-Experts architecture that only partially mirrors the DeepSeek and GLM designs that other Chinese labs have converged on. The backbone consists of 78 layers. The first layer uses a standard dense feedforward network, while the remaining 77 layers replace dense computation with MoE blocks, each containing 256 routed experts and one shared expert. Every token activates the top-8 routed experts along with the shared expert.
The model carries 770 billion total parameters, of which 49 billion are activated per token. A native MTP (multi-token prediction) layer adds another 10 billion parameters (0.7 billion activated) for speculative decoding. The hidden size is 6144, with 64 attention heads using Gated DeepSeek Sparse Attention (Gated DSA) combined with IndexCache for cross-layer sparse index reuse.
The attention module uses a query compression dimension of 2048 and a key-value compression dimension of 512. Each indexer head has 128 dimensions and operates at top-k 2048. The residual pathway uses iHC (identity Hyper-Connections) to expand inter-layer information flow, with 4 residual streams. The MoE intermediate size is 2048 and the FFN intermediate size is 18432.
How Hy4 Compares to the Chinese Open-Weight Field
Tencent's newest preview at 770 billion total parameters now sits closer in scale to Zhipu AI's GLM-5.2, which sits at roughly 753 billion total parameters, and considerably below DeepSeek-V4-Pro's reported 1.6 trillion. The previous July 6 Hy3 release was notably smaller at 295 billion total parameters with 21 billion active, plus a 3.8 billion parameter MTP layer.
Tencent's earlier Hunyuan-Large from October 2024 was described at the time as the largest open-source Transformer-based MoE model in the industry, with 389 billion total and 52 billion active parameters. The July 6 Hy3 release used Apache 2.0 and posted on Hugging Face, ModelScope, GitCode, and CNB simultaneously. The August 28 Hy4 preview appears to follow the same multi-platform distribution pattern.
The naming across the Hy3 and Hy4 lineages introduces some ambiguity. Multiple outlets initially misidentified the August 28 model as Hy3 when it is actually Hy4, reflecting the rapid cadence of Tencent's releases. Tencent's own Hugging Face page uses "Hy4 preview" consistently, and the model card references a progression from Hy3 to Hy4 rather than a revision of the same model.
Training Data and Real-World Use Cases
According to Tencent's model card, Hy4 preview was trained using data built around the actual work performed by internal experts at Tencent, including software engineers, game developers, finance analysts, and security researchers. The model card explicitly lists four application areas:
Software engineering improvements focus on understanding, planning, debugging, and verifying long-horizon development tasks, with gains in front-end visual taste and interaction quality. Office and analysis capabilities convert messy context across multiple files into shareable artifacts including documents, spreadsheets, and presentations, handling data analysis, equations, and financial models with greater precision. Game development capabilities turn a single prompt into a playable prototype while working fluently with game engines for multi-turn refinement. Scientific research applications target AI research, molecular dynamics, condensed matter physics, and pure mathematics.
Tencent partnered 163 internal experts to run a blind side-by-side evaluation on 203 engineering tasks. Hy4 preview scored 2.99 out of 3.0 against GLM 5.3 at 2.92 (46.8 percent wins, 12.8 percent ties, 40.4 percent losses) and against Kimi K3 at 2.94 (51.2 percent wins, 7.9 percent ties, 40.9 percent losses). The evaluation methodology and raw data are not publicly available beyond the published scores.
CodeBuddy and WorkBuddy Integration
Tencent plans to integrate Hy4 preview into CodeBuddy and WorkBuddy, its coding assistant and workplace productivity products. This mirrors a commercial pattern Tencent has used since the July Hy3 launch, when it offered free API access through OpenRouter for a two-week window before shifting to paid API pricing. The promotional free period followed by metered pricing drives fast adoption and benchmark testing in the first two weeks, then converts a portion of that traffic into paying API customers.
The model card references two arXiv papers (2512.02556 and 2603.12201) that likely contain the technical details behind the Gated DSA attention and the MTP speculative decoding layer. These papers have not yet been independently reviewed or replicated, and the model card itself acknowledges "real headroom left in both pre-training and post-training."
Known Limitations
Tencent's model card openly lists several known issues with the Hy4 preview release. The model spends longer than necessary reasoning through complex tasks and shows a tendency to over-verify its own work. The company's approach follows a pattern established with Hy3: ship early, gather feedback, and iterate quickly. The model card states that this early release approach "is what made Hy3 substantially better, and it's how we will get Hy4 right."
The 1 million token context length is impressive on paper, but the practical utility of such a large window depends on effective attention mechanisms. Gated DeepSeek Sparse Attention with IndexCache is designed to manage this efficiently, but real-world performance with documents exceeding hundreds of thousands of tokens will only be clear once independent researchers test the model at scale.
Tencent also released a quantized variant, Hy3-FP8, aimed at teams that want to run the model with a smaller memory footprint. The FP8 variant was posted to Hugging Face alongside the full-precision weights.
What This Means for Open-Source AI
The Hy4 preview release reinforces a broader trend among Chinese AI labs: using open-weight models as both a technical proving ground and a commercial distribution channel. Tencent open-sources the weights to build developer mindshare, establish benchmark credibility, and drive adoption of its paid API and enterprise products. The Apache 2.0 license ensures developers can freely fine-tune, redistribute, and embed the model without the restrictions that some other Chinese open-weight licenses impose.
For the open-source AI community, Hy4 preview adds another high-parameter model to the pool available for research, fine-tuning, and deployment. The 770 billion parameter scale puts it in direct competition with DeepSeek-V4-Pro and Zhipu AI's GLM line for the title of largest usable open-weight model. How it performs under independent benchmarking will determine whether its parameter count translates to real capability gains or remains a marketing figure.
For more on open-weight AI model releases, see our DeepSeek V4: Open Weights, Agent Harness, and Multimodal guide. For context on how open-source AI platforms are being acquired by infrastructure companies, read our Nvidia to Acquire Hugging Face for $12.9 Billion analysis.