Tencent released Hy4 Preview, a large language model whose weights anyone can download. It handles text only and does not accept images. The model is published on Hugging Face. It is far larger than Hy3, the company's July release, and holds much more text in view at once. Its chat template offers only two reasoning settings: "high", the default, and "no_think". Simon Willison tested it through OpenRouter with his standard drawing prompt and published the result. He noted that the model's visible reasoning notes use clipped, ungrammatical English. He suggests that tidy grammar is not worth the extra words in hidden reasoning text.
What changed
Tencent's previous open model Hy3, from July, had 295B parameters and a 256,000-token context.
What it unlocks
Running or calling a freely downloadable Tencent model that holds about a million tokens of text at once.
- 770B total, 49B active parameters
- 1M token context window
- 1.56TB download size
- previous Hy3: 295B, 256K context
What you need to act on it
- download of 1.56TB of weights and substantial hardware, or access through a hosting provider such as OpenRouter
Sources