vram.run Models Hardware Providers Cloud State of Inference
API provider data is live · Hardware & cloud pricing curated 2026-02-23

ERNIE-4.5-21B-A3B-Thinking

by baidu

21.8B params · text-generation · 774 likes · 2.0k downloads

ERNIE-4.5-21B-A3B-Thinking is a 21.8B parameter model. At Q4 quantization it requires 11GB of VRAM. It runs comfortably on GeForce RTX 4090 (60 tok/s), GeForce RTX 5090 (90 tok/s), A100 PCIe 80 GB (110 tok/s).

GPU compatibility

GPUVRAMQ4 DecodeVerdict
GeForce RTX 409024GB60 tok/scomfortable
GeForce RTX 509032GB90 tok/scomfortable
M4 Max 128GB128GB22 tok/stight
M4 Pro 48GB48GB11 tok/stight
M4 Pro 24GB24GB11 tok/stight
A100 PCIe 80 GB80GB110 tok/scomfortable
H100 SXM5 80 GB80GB221 tok/scomfortable
GeForce RTX 309024GB53 tok/scomfortable
Radeon RX 7900 XTX24GB43 tok/scomfortable
GeForce RTX 408016GB42 tok/scomfortable
Install CLI [email protected] Raw data · MIT · API data: live · HW/Cloud data: curated 2026-02-23 · v0.6.0