Saturday, November 21, 2026 · Hy4

From Issue 109 (2026-11-21) · 12 stories in this issue

06 MODEL

❯ Tencent releases weights for its 770-billion-parameter Hy4 flagship with a context window above one million tokens

The November 4 flagship gets downloadable weightsTencent published Hy4 weights on GitHub, Hugging Face, ModelScope and GitCode on November 20. It debuted the production model on November 4. Developers can now download and deploy it themselves, joining publicly available weights for Hy3 preview, Hy3, Hy4 preview and Hy4 flash.

Flagship and Flash deployment costsHy4 has 770 billion total parameters, activates about 49 billion per token and supports more than one million tokens of context. The previously released Hy4 flash has 224 billion total parameters, including 64 billion in its Engram component, activates about 12 billion per token, and supports a million-token context and image-text inputs. Active parameters determine how much of the model each step uses. Deployment also depends on where all the weights are stored and how much cache a long context requires.

Cache, licensing and evaluationTencent says Flash incorporates hybrid attention, LatentMoE and Engram, giving it a key-value cache about one-sixth the size of the flagship’s. Hy4 and Hy4 preview are available under Apache 2.0, which permits commercial use and modification. Tencent also released a system card covering the model’s safety evaluations.

▮ SIGNALWith the flagship weights available, deployment teams can compare Hy4 and Flash directly on long-document work, model storage and cache cost.