Tencent Hy4: a 770B model that can see the screen and drive a computer, with weights promised soon
What is it?
Hy4 is the flagship language model from Tencent's Hunyuan team. It builds on the August preview (a mixture-of-experts model, meaning only part of it runs for each token, with 770B total and 49B active parameters and a 1M-token context) with more pre-training and post-training. This release adds image understanding and computer use for the first time. It can mix command-line and graphical interfaces, and Tencent says it trained specifically on professional software such as Blender and CAD. You can call it as tencent/hy4 on Tencent Cloud and OpenRouter.
Why does it matter?
Few models with promised open weights handle both vision and computer use. Pricing matches the preview, and OpenRouter lists it at $0.417 input and $1.251 output per million tokensVendor claim with a 50% discount applied. Tencent says its own search evals show the hallucination rate falling from 15.76% to 5.45%Vendor claim, and its GameDev score rising 19.1%Vendor claim. That makes it worth trying as a cheap engine for document, finance, or game-development agents.
Maturity
GAThe API is generally available. The weights and system card are only promised "soon," with no date or license yet, and every performance number comes from Tencent's own testing.
How to track it
- Weights: Watch the tencent account on Hugging Face for when the Hy4 weights and system card appear, and under which license.
- Hy4 flash: Check the hy.tencent.com research page for the cheaper, faster Hy4 flash that Tencent has announced.
- Independent tests: See whether outside benchmarks such as Artificial Analysis back up the computer-use and hallucination numbers.