Tencent Hunyuan publishes the Hy4 Flash repository and links to GGUF weights
Tencent Hunyuan published the Hy4 Flash repository, documenting a model with 220B total parameters, about 12B active per token, and a context window of up to 1M tokens. It accepts text and images and generates text. The repository links to a public GGUF model page on Hugging Face.
China context
- Original name
- 腾讯
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Use the linked GGUF weights to test self-hosting, and measure memory use at your target context length before selecting hardware.
Tencent Hunyuan's GitHub repository describes Hy4 Flash's mixture-of-experts architecture and documents deployment paths for Transformers, vLLM, and SGLang. It also reports base-model benchmarks for text and image tasks.
Hy4 Flash combines sparse attention with sliding-window attention and uses a built-in DFly2 drafter that reuses the model's attention cache during speculative decoding. The active parameter count does not establish how much memory a deployment needs. A useful independent test would report peak memory and generation throughput at several context lengths on named hardware.
Developers outside China can access Hy4 Flash through the linked GGUF weights without relying on a Tencent-hosted API. The repository does not establish hosted API availability outside China or a cost advantage over other models. Comparable deployment tests would show whether the model is cheaper to run at similar accuracy and latency.
Infrastructure teams can use the public weights to test Hy4 Flash on their own text and image workloads. A purchasing decision still needs measured cost per completed task, including hardware memory requirements and latency at the team's typical input lengths.
The next checks are independent runs of the linked GGUF weights at long context lengths and a public Tencent API listing that states supported regions and pricing.