Tencent just dropped a preview of Hy4, a new open-weight large language model, and the numbers are the headline. According to Simon Willison, who flagged the release on his link blog, Hy4 clocks in at 770 billion total parameters with 49 billion active, a 1 million token context window, and a 1.56TB footprint on Hugging Face. It’s a text-only model for now, with no vision capabilities.
What stands out is the jump in scale. Willison notes that Hy4 more than doubles the company’s previous release, Hy3, which shipped back in July at 295B total parameters, 21B active, a 256,000 token context, and 598GB on disk. In roughly a month, Tencent has pushed its open-weight lineup into serious frontier territory.
What’s new in Hy4
Here’s what the preview brings, based on Willison’s breakdown:
- 770B total parameters, 49B active. It’s a mixture-of-experts design, so only a fraction of the weights fire on any given token. That keeps inference costs lower than the raw parameter count suggests.
- 1 million token context window. A 4x expansion over Hy3’s 256,000, which opens the door to long documents, large codebases, and extended conversations.
- Text input only. No image or vision support in this release.
- Open weights on Hugging Face. The full 1.56TB model is available to download, which puts it in reach of anyone with the hardware to run it.
Two reasoning modes, not a slider
Willison has been reading model chat templates lately to understand how these systems actually behave, and Hy4’s template revealed something specific. The model exposes just two reasoning effort levels: “high” (the default) and “no_think” (reasoning disabled). Pass anything else and the template raises an error.
That’s a simpler setup than the graduated reasoning controls some competitors offer. You either get full deliberation or none. For developers, it means less tuning and a clearer mental model of what the system is doing on each request.
The pelican test
Willison ran his signature benchmark, “Generate an SVG of a pelican riding a bicycle,” through Hy4 with default high reasoning via OpenRouter. The output itself was one data point, but the reasoning trace was the more interesting part.
The model visibly debated design choices: “Let’s maybe add a helmet? It could improve riding theme, but may obscure head… But pelican with big beak; a helmet might obscure. Better maybe no.” It cycled through sunglasses and water before dropping both.
Willison points out that the trace reads in slightly broken English. His take is that perfect grammar isn’t useful or token-efficient for hidden reasoning text, so the model doesn’t bother. It’s a small window into how these systems economize when they think nobody’s reading.
Why it matters
Hy4 is another sign that the open-weight race is being run hard from China. Tencent is iterating fast, roughly a month between major releases, and pushing scale aggressively. A 1M context window and 770B parameters on freely downloadable weights is the kind of spec sheet that used to belong only to closed, API-gated frontier models.
The MoE architecture matters here too. With only 49B active parameters, teams don’t need to serve the full 770B on every call, which makes running a model this large more practical than the headline number implies. Still, 1.56TB of weights is not a casual download, so this is aimed at organizations with real infrastructure rather than hobbyists on a single GPU.
A few caveats worth keeping in mind. This is a preview, not a final release, so behavior and benchmarks may shift. The text-only limitation rules out multimodal use cases. And Willison’s reporting focuses on specs and early impressions rather than head-to-head evals against models like DeepSeek or Qwen, so how Hy4 actually stacks up on quality is still an open question.
For now, the trajectory is the story. Tencent is scaling its open models faster than most expected, and the gap between what’s free to download and what’s locked behind an API keeps narrowing. You can find the full write-up and the pelican SVG at the original source.