Key takeaways
- Tencent published Hy4 flash weights on Hugging Face in three versions: chat, FP8 and base, under the Tencent Hy License.
- The mixture-of-experts model has 220 billion parameters, about 12 billion active per token, and a 1,048,576-token context.
- Tencent says a task costs one eighth of what it does on DeepSeek-V4.1-Flash, by its own measure.
Tencent has put the weights of Hy4 flash, the cheaper version of its Hy4 model, on Hugging Face for anyone to download. There are three versions: the chat model, an FP8 copy that needs less memory, and the base model before chat training. All three come under the Tencent Hy License.
When Tencent released Hy4 in early November, it said a Flash version was coming but gave no date. Its third-quarter results, out on November 18, say Hy4 flash "brings down inference costs", and the same report shows the company spending RMB81.4 billion in a quarter on compute and data centers.
Hy4 flash is a mixture-of-experts model, which runs only part of its network for each token. It has 220 billion parameters in total and about 12 billion active per token. It reads text and images, writes text, and holds up to 1,048,576 tokens of context. A small built-in drafter model guesses the next tokens ahead, which Tencent says helps push decoding past 150 tokens a second.
The numbers on the model card come from Tencent. It says a task costs about one thirteenth of what it did on the Hy4 preview and one eighth of DeepSeek-V4.1-Flash. In a blind test with 85 users and 259 tasks, people rated it 2.79 out of 4, against 2.81 for Claude Haiku 5.5. None of these scores has been checked independently yet.
It is not a laptop model. Tencent's vLLM example splits it across eight GPUs, and Transformers support still needs an unmerged pull request. Docker images for vLLM and SGLang are on the model page.