Tencent released Hy4 Preview on August 29, an open weight, text-only large language model with 770 billion total parameters, 49 billion of them active, and a 1 million token context window. The download on Hugging Face runs 1.56 terabytes. That figure, more than the parameter count, decides who actually gets to run this model: a lab with a multi-GPU cluster can load it, an individual developer on a single consumer card cannot.
The jump from Tencent’s prior release is steep. Hy3, published in July, shipped at 295 billion total parameters, 21 billion active, a 256,000 token context window, and a 598 gigabyte weights file. Hy4 more than doubles the active parameter count, quadruples the context window, and nearly triples the file size in one generation.
The gap between Hy4’s 770 billion total parameters and its 49 billion active parameters points to a mixture of experts design, where only a portion of the network runs on any given token. That architecture keeps inference cost closer to a 49 billion parameter model even as the published weights file balloons. It does not change the storage bill. Someone still has to hold 1.56 terabytes on fast disk or across GPU memory before the model produces a single token, and that constraint separates a well-funded lab’s version of “open weights” from an individual developer’s.
Simon Willison, the independent developer who maintains Datasette, examined Hy4’s published chat template and found it supports exactly two reasoning settings: “high,” the default, and “no_think,” which disables reasoning outright. Writing on his blog on August 29, Willison tested the high-reasoning setting through OpenRouter using his own “pelican riding a bicycle” SVG-generation prompt. He noted that the hidden reasoning trace read in deliberately truncated, ungrammatical English, weighing and then rejecting small additions like a helmet or sunglasses for the pelican. Willison’s read is that clean grammar costs tokens the model has no reason to spend on text nobody is meant to see, a small but concrete data point on how these labs optimize hidden reasoning differently from the polished answer a user actually reads.
Hy4 is text input only, with no vision component, a scope Tencent stated plainly. That rules out image understanding tasks and narrows the model’s competitive set to other text-only open weight releases.
Neither Tencent’s release notes nor Willison’s testing includes benchmark scores against other open weight systems near Hy4’s size. The only comparison on record is Tencent’s own prior model, not a third-party evaluation, so how Hy4 performs against the rest of the open weight field is still unverified.
For teams evaluating open weight infrastructure, the 1.56 terabyte footprint is the real gating question, ahead of parameter count or context window. Confirm the model fits the inference stack you actually have before you spend time benchmarking what it can do.
Simon Willison reported the Hy4 Preview release and its reasoning-template details on his blog on August 29, 2026.
