Tencent Hunyuan announced on November 18 that the weights of Hy4 Flash are now open-source: the standard build, an FP8 quantized build, and a Base build are all available on GitHub, Hugging Face, ModelScope, and GitCode. It is the second move in two days since Hy4 Flash launched as an API-only model on November 16: first sell the service, then release the weights. The open-source path for Tencent's Hy family is accelerating.
What exactly was open-sourced
Hy4 Flash is the lightweight version of Hy4: 224 billion total parameters, 12 billion activated per inference, a 1-million-token multimodal context window, aimed at everyday office work. When it launched on November 16 it was API-only, priced at 0.5 yuan per million input tokens and 2 yuan per million output tokens — about $0.07 and $0.28 — with a half-price launch offer and free access in WorkBuddy until November 30. With the weights open, developers can download and deploy the model themselves instead of relying on Tencent Cloud's API. Pricing at the November 16 launch had already cut costs to one-twentieth of full Hy4, and open weights let teams do the math again: those with their own servers and GPUs can keep inference costs in their own hands. The three builds have distinct roles: the standard build for full deployment, the FP8 quantized build for lower memory use, and the Base build for teams that want to fine-tune it themselves. License terms follow whatever each repository page states — check before downloading.
Low cost is the selling point, but the barrier hasn't vanished
Tencent's internal WorkIndex benchmark shows Hy4 Flash completing office tasks at one-twentieth the cost of full Hy4; in a blind test where 85 WorkBuddy users rated 259 tasks, it scored 2.79 out of 4, close to Claude Haiku 5.5's 2.81 and above GLM-5.3-Flash's 2.53. But these are Tencent's own numbers, and independent evaluations have not caught up yet. The more practical barrier is hardware: a 224-billion-parameter model, even in FP8, does not fit on an ordinary graphics card, so individual developers will most likely still find the API the better deal. The real beneficiaries of open weights are mid-to-large teams with bulk inference needs and vendors that want to build on top of the model. In other words, open weights expand what Hy4 Flash can be used for; they do not turn it into a local model anyone can run. Questions such as what happens to WorkBuddy chats after switching computers will not change because of open weights: model weights and product data are two different things.
For Tencent, the intent is clear: full Hy4 launched on November 4 as API-only, with its weights still nowhere to be seen; open-sourcing the lightweight version first answers the open-source community's expectations and drives traffic to in-house products such as WorkBuddy and Yuanbao. What to watch next is whether full Hy4's weights follow, and whether the open-source community builds interesting fine-tunes on the Flash version. Until the free period ends on November 30, ordinary users can try it directly in WorkBuddy without waiting for deployment guides.