LimitDigest

Hy4 Flash VRAM requirements

Tencent · 220B params (12B active) · Tencent Hy License. Estimates, not vendor numbers.

Quick answer

Hy4 Flash (220B parameters, 12B active per token) needs about 134 GB of VRAM or unified memory at Q4, 158 GB at Q5 and 235 GB at Q8 (weights plus about 1.5 GB headroom, our estimate). It is too large for the consumer setups we list; it needs a multi-GPU rig or a very large-memory Mac or server.

QuantDownloadMemory needed
Q4132 GB134 GB
Q5157 GB158 GB
Q8234 GB235 GB

Will it run on my machine?

SetupRuns?Speed est.
8GB GPU (RTX 4060)Non/a
12GB GPU (RTX 3060 / 4070)Non/a
16GB GPU (RTX 4060 Ti 16GB)Non/a
24GB GPU (RTX 4090)Non/a
32GB GPU (RTX 5090)Non/a
Mac 16GB (M4)Non/a
Mac 48GB (M4 Pro)Non/a
Mac 128GB (M4 Max)Non/a
PC, 32GB RAM, no GPUNon/a

Best quant that fits each setup, with estimated generation speed from memory bandwidth. Real speed depends on the runtime, context length and settings.

Mixture of experts: 220B total (64B of it Engram tables), 12B active per token, 1M context, text and image input. Open weights released Nov 17, 2026; the FP8 build is about 287 GB, so it needs a very large Mac or multi-GPU box.

FAQ

How much VRAM does Hy4 Flash need?

About 134 GB at Q4, 158 GB at Q5 and 235 GB at Q8, including about 1.5 GB headroom. These are estimates from the parameter count on the official model card.

Can Hy4 Flash run on 8GB of VRAM?

Not fully on the GPU. At Q4 it needs about 134 GB; partial CPU offload works but is much slower.

Can Hy4 Flash run on a Mac?

Not on the Mac configurations we list; it needs more unified memory.

Which quant of Hy4 Flash should I use?

Use the highest quant that fits with room to spare: Q8 is near full quality, Q5 is a good balance, and Q4 is the smallest that usually keeps quality acceptable.

Why is Hy4 Flash fast for its size?

It is a mixture of experts: 220B parameters in total but only 12B are used per token, so it reads less memory per token. You still need memory for all 220B.

Similar-size models

Recursive Flint 40B A2.5B (39.8B) · Llama 3.3 70B Instruct (70B) · gpt-oss-120b (117B) · Mistral Small 4 (119B) · Ornith 1.5 122B (122B) · Qwen3.8 Flash Next (180B) · Muse Spark Open (816B)