WHAT SHIPPED
● Tencent Hy4 Preview: 770 billion total parameters, 49 billion active, 1M context window.
● Licence: Apache 2.0. Not custom, not modified, not unstated.
● Why that is notable: it is the first open-weight flagship this summer to use a standard permissive licence.
The licence is the story
We wrote last week that four open-weight releases this summer shipped under four different licences and not one was Apache. Qwen 3.8-Max used custom terms. Kimi K3 used a modified MIT. Z.ai held GLM-5.3 weights for two weeks and has still not stated the flagship licence.
Hy4 Preview breaks that. Apache 2.0 is a standard, well-understood licence with no modifications to read, and applying it to a 770-billion-parameter flagship is a meaningful choice rather than a default.
WHY A STANDARD LICENCE IS WORTH MORE THAN IT SOUNDS
A modified licence has to be read clause by clause, and the modification is where the restriction lives. A custom licence has no case law and no shared understanding behind it.
Apache 2.0 means your lawyer has seen it before, which for a company deciding what to build on is worth more than a few benchmark points.
The architecture
| Property |
Hy4 Preview |
| Total parameters | 770 billion |
| Active per token | 49 billion — roughly 16 to 1 sparsity |
| Context window | 1 million tokens |
| Licence | Apache 2.0 |
| Status | Preview |
The sparsity ratio sits close to GLM-5.3's roughly nineteen to one and well below Qwen3.8-Flash-Next's twenty to one, so the architecture is conventional for a 2026 mixture-of-experts flagship. Nothing here is an architectural claim — the licence is the differentiator.
What it does not tell you
- Benchmarks. No independent evaluation exists yet. Preview releases rarely arrive with one.
- Hardware requirements. At 770 billion parameters this is a multi-node deployment, not something you run on a workstation.
- Training data provenance. Apache 2.0 governs what you may do with the weights. It says nothing about where the training corpus came from, which is a separate exposure and the one currently in front of courts.
Where this leaves the open-weight field
| Release |
Terms |
| Tencent Hy4 Preview | Apache 2.0 |
| K2 Horizon | Plus training data, code, checkpoints and logs |
| GLM-5.3-Flash | MIT |
| Kimi K3 | Modified MIT — read the modification |
| Qwen 3.8-Max | Custom licence |
| GLM-5.3 flagship | Still unstated |
Two of the six now meet what open source is normally taken to mean. A month ago it was arguably none, so the direction is better than last week's piece implied — worth saying, since we wrote that one.
Sources
FAQ
What is Tencent Hy4 Preview?
An open-weight mixture-of-experts model at 770 billion total parameters with 49 billion active per token and a 1 million token context window, released under Apache 2.0.
Why does Apache 2.0 matter here?
Because four earlier open-weight flagships this summer used custom, modified or unstated terms. Apache is standard, unmodified and familiar to legal teams, which lowers the cost of building on it.
Is it better than GLM-5.3 or Kimi K3?
No independent benchmarks exist yet. The differentiator today is the licence, not measured capability.
Can I run it locally?
Not on a workstation. At 770 billion parameters this is a multi-node deployment.
Does Apache 2.0 cover the training data?
No. It governs what you may do with the weights. Where the training corpus came from is a separate question, and the one currently before courts.