Radar· local AI & agent security
models high confidence

Hy4 Flash goes open-weight: 224B (12B active), 1M context and an official GGUF

Nov 18, 2026 · tencent, hy4, moe, gguf

TL;DR Tencent opened Hy4 Flash, its FP8 build and the Base model. Official GGUF from 126 to 157 GiB; the 59 GiB Engram table can live in RAM or on NVMe. Custom license.

In plain words: a large Chinese AI model is now downloadable. It is a “mixture of experts”: huge in total, but only a small slice (12B parameters) works on each word, which is what makes it runnable on strong home hardware. Part of it is a lookup table that does not need to sit on the graphics card.

What happened

Tencent published Hy4 Flash, Hy4 Flash-FP8 and Hy4 Flash-Base on Hugging Face and GitHub:

  • 224B parameters in total, of which 64B are an Engram table;
  • 12B active per token;
  • 1M-token context, text and image input.

The full Hy4 comes next, according to Tencent.

Who is affected

People running large models locally with a lot of RAM or fast NVMe plus one or more big GPUs.

Why it matters

The official GGUF (AngelSlim) ships IQ2_AH at 125.7 GiB, Q3_K at 144.7 GiB, Q4_K_M at 156.8 GiB and SPARK-4Bit builds of about 150 GiB. The 59 GiB shard is the Engram table, which can stay in RAM or be memory-mapped from NVMe, leaving roughly 155B for the GPU side. Tencent’s own number: 48 tok/s on a 128 GB Mac M5 Max with Q4_K_M at 4K context (vendor claim).

What to do

  1. llama.cpp support is still a patch published by Tencent, not merged upstream. Expect rough edges.
  2. The license is the Tencent Hy License, Apache-derived with extra terms. Reports say personal, research and commercial use are allowed, while large API services need a separate license. Read the text before you use it.
  3. As with any download: pin the revision and check the SHA-256 of each file.

Sources

← back