AI Daily · 2026-08-30

Tencent released Hy4 Preview, a text-only open-weight model that significantly outscales July's Hy3 with 770B total parameters, 49B active parameters, a 1M-token context window, and roughly 1.56TB of weights on Hugging Face. Simon Willison found via its config that it only offers two inference modes: high and no_think. When he asked it to generate an SVG of a pelican riding a bicycle, the hidden reasoning trace showed the model debating whether to add a helmet, sunglasses, and water before deciding against all three. He also noticed the hidden reasoning text uses truncated English and speculated that full grammar adds little to reasoning while consuming extra tokens. Overall, this marks another escalation in the scale and context race among open-weight Chinese models.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐ [Model Release] Introducing Hy4 Preview

Tencent released Hy4 Preview, a new open-weight text-only LLM with 770B total parameters, 49B active parameters, a 1M-token context window, and about 1.56TB of weights on Hugging Face—a large increase from July's Hy3 (295B total, 21B active, 256K context). Simon Willison inspected Hy4's chat template and found only two reasoning-effort settings: the default 'high' and 'no_think', which disables reasoning. Testing with an SVG-generation prompt under the default high mode, the model's reasoning trace weighed whether to add a helmet, sunglasses, or water to a pelican riding a bicycle, then decided not to. He also observed that the hidden reasoning text uses truncated English, presumably because perfect grammar is neither useful nor token-efficient for internal reasoning.