Tencent Hunyuan released and open-sourced its new large language model, Hy4 preview, on Aug. 28. The model has 770 billion total parameters, activates 49 billion parameters per token, and extends context length beyond 1 million tokens. Tencent has published the model weights on Hugging Face under the Apache 2.0 license.
According to Xinhua, Tencent describes Hy4 preview as a model "built for productivity," with target use cases in software engineering, office analysis, game development, and scientific research. The release is another open-source push from a major Chinese technology company after Zhipu’s GLM-5.3 and the DeepSeek V4 series.
78-layer MoE design
Based on the model card, Hy4 preview uses a mixture-of-experts architecture with 78 layers. The first layer is a dense feed-forward network, while the remaining 77 layers are MoE layers. Each layer includes 256 routed experts and one shared expert, and each token selects the top eight routed experts.
The model includes a multi-token prediction layer to support speculative decoding, and Tencent also provides an FP8 quantized version to reduce memory requirements. A deployment example uses eight GPUs for tensor parallelism. Tencent notes in the model card that Hy4 preview is still an early version of Hy4 and that inference efficiency has known limitations.
Benchmarks and internal blind testing
The model card lists the following scores: 92.3 on GPQA Diamond, 82.9 on SWE-bench Multilingual, 65.7 on SWE-bench Pro, and 85.4 on Terminal-Bench 2.1.
Xinhua also cited Tencent’s internal evaluation. In a blind test covering 203 engineering tasks, 163 internal experts gave Hy4 preview an average score of 2.99 out of 4. That was slightly above GLM 5.3 at 2.92 and Kimi K3 at 2.94.
All of those figures were disclosed by Tencent, and independent third-party verification has not yet been completed.
Used in its own optimization work
The report also says Hy4 preview was involved in its own optimization process, including training methods, data strategy, evaluation systems, and inference infrastructure. Tencent said end-to-end inference throughput improved by 31.8%.
Chain News previously reported that Zhipu launched GLM-5.3-Flash at the end of August for operation on domestic chips, while DeepSeek open-sourced V4.1-Flash on Sept. 10. The three companies updated their models within a two-week span.
Pricing and availability
Xinhua reported that the API is priced at RMB 6 per million input tokens, RMB 18 per million output tokens, and RMB 0.3 for cache hits. The model is available through Tencent Cloud Tokenhub and OpenRouter, and it has also been integrated into Tencent’s WorkBuddy and CodeBuddy.
Using the Sept. 11 exchange rate cited in the report, input pricing works out to about $0.83 and output pricing to about $2.5. The report says that places Hy4 preview between the off-peak pricing of DeepSeek V4.1-Flash and U.S. closed-source models.