Tencent Hy4 flagship weights are now public, with a catch for home hardware

Text
iHOLYSHITVISUAL / TECHModel releasesiHOLYSHIT / ILLUSTRATION
iHolyShit illustration

Tencent has released downloadable weights for its flagship Hy4, including a lower-precision version and the base model. The flagship was already accessible through an API. What changes now is who can test the model on their own hardware, and on what terms.

On November 20, Tencent put the weights for its flagship Tencent Hy4 on Hugging Face. The public repositories contain an instruction-tuned model, a lower-precision variant and the base checkpoint from before chat tuning. These serve different purposes: testing the finished model, reducing storage needs or studying what existed before chat tuning. An API lets you send a task to someone else’s computer. Downloadable weights let teams with the resources choose their own test environment.

This is not the flagship’s debut

This is not the flagship’s debut. An API was available on November 4; Tencent Hy’s November 18 announcement covered downloadable weights for the lighter Flash model. An earlier preview of comparable scale also had downloadable weights. This release makes three checkpoints in the flagship family available. On November 19, Together AI showed its hosted model reading a chart and recreating it as a web page. That is one task demonstrated by a hosting provider, not a broad independent assessment of the released weights.

The Tencent Hy team’s November 20 Hugging Face model card lists 770 billion total parameters and about 49 billion active per token in the language backbone, with support for up to one million tokens of context. Think of a workshop where only some benches are used for a given job. The other benches are still there for the next job. Likewise, activating fewer parameters does not cut the storage needed for all the weights by the same proportion. The ability to accept long inputs does not guarantee reliable recall of a particular detail.

The license depends on the version

The license depends on the version. The preview used Apache 2.0; the flagship weights use Tencent Hy License 1.0. It sets additional licensing conditions for large operators that offer model access to third parties, not a blanket ban on commercial use. Small teams may first need to decide whether maintaining their own environment is worth it. For teams evaluating the model, the weights add an option: rerun the same material in different environments rather than rely on a vendor demo. Whether it can run and whether it can do the work reliably are separate questions.

iHolyshit review I’m more interested in whether teams can turn downloadable weights into repeatable work than in another leaderboard screenshot. Over the next six to twelve months, I’ll look for public comparisons that reuse the same chart images, track whether the numbers and units survive, retain failures and report time and cost by version. That would tell me more about where to use the flagship than one impressive reconstruction.

Share this article

What do you think of this story?

Anonymous feedback, up to 5 reactions per article.

Reader comments

Comment without signing in, or sign in to display your member name.

Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.

0/500

No comments yet. Be the first to share.

Sources