↓Skip to main content
Zeitgeist — a spike by Chris Gathercole
  1. Creators/
  2. Simon Willison — Creator of Datasette; co-creator of Django/

Introducing Hy4 Preview

Blog · Watch/Read · (gathered 2026-09-03)

  • Simon introduces Tencent’s new Hy4 Preview LLM, highlighting its massive scale with 770B total parameters, 49B active parameters, and a 1M token context window, a significant jump from Hy3.
  • He specifically investigates Hy4’s chat_template.jinja to uncover its internal “reasoning effort” levels, which are defined as “high” (the default) and “no_think.”
  • A test prompt (“Generate an SVG of a pelican riding a bicycle”) reveals the model’s internal “reasoning trace,” demonstrating a deliberative process of considering and rejecting ideas (e.g., adding a helmet or sunglasses).
  • The “reasoning trace” uses “slightly truncated English,” suggesting an optimization for token efficiency in the model’s hidden internal thought processes rather than perfect grammar.