Tencent's Hy4 model runs 770B params, 49B active

Tencent's Hy4 preview packs 770 billion total parameters, 49 billion active, and a 1 million token context window, per Simon Willison's testing. The full weights come to 1.56TB. That size puts self-hosting out of reach for most local setups.

Hy3, the prior version, ran 295 billion total parameters with 21 billion active and a 256K context window. Hy4 also ships two reasoning effort levels, with "high" as the default and a "no_think" mode for skipping the extra deliberation.

Willison's test was an SVG generation task. The reasoning trace showed the model weighing specific details, considering a helmet, considering sunglasses, then rejecting both. It read in slightly truncated English, more like notes than full sentences.

Willison's writeup doesn't include a benchmark comparison against Hy3 or against other large models at similar active-parameter counts. He reported what he saw running it himself. He didn't say whether the extra size or the larger context window improves output quality beyond that one SVG test.