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

Kimi-Linear-48B-A3B-Instruct

by moonshotai

49.1B params · text-generation · 549 likes · 40.1k downloads

Kimi-Linear-48B-A3B-Instruct is a 49.1B parameter model. At Q4 quantization it requires 25GB of VRAM. It runs comfortably on GeForce RTX 5090 (40 tok/s), A100 PCIe 80 GB (48 tok/s), H100 SXM5 80 GB (98 tok/s).

GPU compatibility

GPUVRAMQ4 DecodeVerdict
GeForce RTX 509032GB40 tok/scomfortable
M4 Max 128GB128GB9 tok/stight
M4 Pro 48GB48GB4 tok/stight
A100 PCIe 80 GB80GB48 tok/scomfortable
H100 SXM5 80 GB80GB98 tok/scomfortable
M4 Max 64GB64GB9 tok/stight
H200 SXM 141 GB141GB143 tok/scomfortable
M2 Ultra 192GB192GB13 tok/stight
Install CLI [email protected] Raw data · MIT · API data: live · HW/Cloud data: curated 2026-02-23 · v0.6.0