Tencent releases Hy4 flash weights for local deployment
Tencent has opened downloads for its Hy4 flash model family, letting developers run the text-and-image model in their own environments. The release includes instruction-tuned and pretrained variants under the Tencent Hy License. Its architecture combines expert layers and hybrid attention, while deployment documentation covers serving tools and fine-tuning. Download access is separate from the existing hosted service.
Artificial Intelligence··Midday
Weights can run in developers’ own environments
Tencent, the technology company behind Hy4 flash, has made the model family’s weights downloadable. Weights are the learned values needed to run a trained model. Developers can now obtain those files and deploy the model in their own environments under the Tencent Hy License. The downloads provide another route alongside the existing hosted service. Tencent’s model page directs developers to model repositories including Hugging Face.[1], [2]
Text and images feed a text-generating model
Hy4 flash takes text and image inputs and returns text, supporting a context window of roughly one million tokens, the units used to process input. Tencent describes an architecture combining mixture-of-experts layers, hybrid attention and learned representations of word sequences. A built-in drafting model shares the main model’s cache and selected positions as part of speculative decoding, a method that drafts possible output in advance.[1]
Chat tuning and deployment follow different paths
Instruction-tuned versions offer high-reasoning and direct-response modes, while Hy4 flash Base is a pretrained model without chat instruction tuning. Deployment instructions cover vLLM and SGLang, tools for serving language models, together with quantization and fine-tuning. Support in Transformers, a model software library, is awaiting merger into its main branch. Tencent directs developers to a specified version until that work is incorporated.[1]