Tencent has put its Hy4 flash model on Hugging Face: 220B parameters, 12B active per token and a 1M-token context, built for cheap and fast AI agents.
Tencent has released the weights of Hy4 flash, a smaller and faster member of its Hy4 model family, on Hugging Face, ModelScope and GitHub this week, a couple of days after making it available through its own apps and cloud API. According to Tencent's model card, Hy4 flash has 220 billion parameters in total, including about 64 billion "Engram" memory parameters, but only around 12 billion are active for each token. It uses a mixture-of-experts design, which means only a small set of specialised sub-networks runs for each request, keeping it fast and cheap. The model reads text and images, writes text and handles up to one million tokens of context. Tencent says it generates more than 150 tokens per second.
Tencent pitches Hy4 flash as a workhorse for everyday agent tasks rather than a rival to its flagship. On its own WorkIndex benchmark, built around tasks in its WorkBuddy assistant, the company reports a score of 42.1, below the full Hy4 model's 51.4, but says each task costs about one twentieth as much as with Hy4 and one eighth as much as with DeepSeek-V4.1-Flash. In a blind test with 85 internal users, Tencent says satisfaction reached 2.79 out of 4, close to the 2.81 it measured for Anthropic's Claude Haiku 5.5. These are company figures that independent evaluators have not yet confirmed. The release adds to a crowded field of Chinese open-weight models from DeepSeek, Alibaba and others competing on price.
Developers can download Hy4 flash for free from Hugging Face under the name tencent/Hy4-flash, with FP8 and base versions also available, and run it with tools such as vLLM and SGLang or Docker Model Runner. It is an open-weight model, meaning the files can be downloaded and run on your own servers, but it ships under Tencent's own Hy License rather than a standard open-source licence, so companies should check the terms before commercial use. Its size still calls for server-class GPUs. Tencent has not published a per-token API price on the model card. Arabic is not among the languages listed in its tests, so performance for users in Morocco and the Arab world remains unclear.



