Introducing Hy4 Preview
Simon Willison’s Weblog Simon Willison ● Covered by 2 sources
Tencent released Hy4 Preview, an open-weight text-only model with a 1M-token context. It’s a huge jump from Hy3, and its hidden reasoning seems to come in only two modes.
Based on reporting by Simon Willison’s Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Tencent has put out Hy4 Preview, a text-only open-weight model with 770B total parameters, 49B active parameters and a 1M token context window. The model is on Hugging Face, and the upload weighs in at 1.56TB. That alone makes it a very large step up from Hy3, which landed in July with 295B total parameters, 21B active parameters, a 256,000 token context window and a 598GB release.
The interesting part is not just the size. Simon Willison dug into Hy4’s chat template on Hugging Face and found a tiny but revealing bit of control flow: if reasoning_effort isn’t set, it defaults to high, and the only accepted values appear to be high and no_think. That suggests Tencent is exposing two modes rather than a whole menu of thinking settings.
Willison also tried a simple image-generation-style prompt — “Generate an SVG of a pelican riding a bicycle” — through OpenRouter with the default high reasoning mode. The reasoning trace he quoted reads like something in the middle of a drafting session, full of quick trade-offs about helmets, caps and sunglasses. The grammar is rough in places, but that may be the point: hidden reasoning text does not need to be polished to be useful.
What stands out here is the mix of scale and restraint. Hy4 is being sold as a giant model, but the chat template suggests a deliberately narrow interface for reasoning. That is a cleaner move than pretending every knob matters. A lot of AI products still confuse more settings with more intelligence.
My take — AI-written commentary, not fact-checked reporting
Tencent is doing the sensible open-model thing here: ship the weights, keep the interface blunt, and let people inspect the guts. The hidden reasoning system sounds less like magic and more like engineering, which is exactly how it should be. The industry could use fewer theatrical “thinking” sliders and more plain defaults that don’t pretend to be a philosophy seminar.
Read more about this at: Simon Willison’s Weblog