
Tencent Puts Hy4 Flash Weights Online, With a Licence Attached

Tencent has published downloadable weights for Hy4 flash, two days after putting the model behind an API. The release covers three checkpoints, the instruction tuned model, a quantised FP8 version, and a pretrained base that is meant for further training rather than chatting.

The change is access rather than new capability. Nothing about the model's specifications or pricing has moved since the API launch, so what a developer gains here is a copy they can run on their own hardware, which is a different proposition from calling someone else's endpoint.
And the sizes explain who this is for. The standard checkpoint's index records roughly 440 gigabytes of tensor data and the FP8 version about 287, and those are weight sizes rather than memory requirements, so the gap between quantised and unquantised is storage rather than a route onto smaller hardware, and a full run still assumes a multi node cluster.
And the deployment paperwork is part of the release. Tencent ships container images for vLLM and SGLang with examples that use eight way tensor parallelism, and its Transformers instructions point at a pending pull request rather than an ordinary package, which tells you the tooling is fresh.
And there is a memory detail worth flagging. The vLLM recipe offloads Engram tables to host RAM by default, so a server needs system memory as well as GPU memory, and the starting configuration uses 8,192 tokens of context even though the model advertises a capacity of one million.
And the licence is where the download stops being simple. The checkpoints use Tencent Hy License 1.0, which says outright that it is not Apache 2.0 despite deriving from it, and section 10 requires a separate written licence for operators running the model as a service above 50 million US dollars in yearly revenue, or in products with more than 100 million monthly users, with a narrow exception for internal use where neither the model nor its outputs reach third parties.
So this is open weight under conditions rather than open source. Tencent says the larger Hy4 model's weights are still being prepared, which makes the flash release a first step rather than the full opening.
Three checkpoints are available, instruct, FP8 and base
The standard weights index at about 440 gigabytes, FP8 at about 287
Deployment needs vLLM or SGLang containers and eight way tensor parallelism
Default serving context is 8,192 tokens, well short of the advertised million
The Tencent Hy Licence adds revenue and user thresholds for commercial serving
Source: Pandaily


