Skip to content

MiniCPM5 2B VRAM requirements

How much VRAM MiniCPM5 2B needs
4-bit weights: 1.4 GBWith 8K context: 2.5 GBWith 32K context: 3.5 GB8-bit, 8K context: 3.6 GB
Parameters
2.5B
Longest context
128K
Released
Sep 2026
License
Apache 2.0
Strongest at
Not rated

How much VRAM MiniCPM5 2B needs

Memory MiniCPM5 2B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit4.7 GB5.8 GB6.8 GB10.7 GB
8-bit2.5 GB3.6 GB4.6 GB8.5 GB
6-bit1.9 GB3.0 GB4.0 GB7.9 GB
5-bit1.7 GB2.8 GB3.7 GB7.7 GB
4-bit1.4 GB2.5 GB3.5 GB7.4 GB
3-bit1.1 GB2.2 GB3.2 GB7.2 GB
2-bit900 MB2.0 GB3.0 GB6.9 GB
1.6-bit495 MB1.6 GB2.6 GB6.5 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run MiniCPM5 2B

The best way MiniCPM5 2B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit104K72 tok/s0.3 sPDF: MiniCPM5 2B on RTX 4060 report
RTX 50608-bit104K113 tok/s0.2 sPDF: MiniCPM5 2B on RTX 5060 report
RTX 30708-bit104K113 tok/s0.2 sPDF: MiniCPM5 2B on RTX 3070 report
RTX 3080 10 GB8-bit128K174 tok/s0.2 sPDF: MiniCPM5 2B on RTX 3080 10 GB report
RTX 3060 12 GB8-bit128K93 tok/s0.4 sPDF: MiniCPM5 2B on RTX 3060 12 GB report
Arc B5808-bit128K77 tok/s0.3 sPDF: MiniCPM5 2B on Arc B580 report
RTX 40708-bit128K124 tok/s0.2 sPDF: MiniCPM5 2B on RTX 4070 report
RTX 4070 Super8-bit128K124 tok/s0.1 sPDF: MiniCPM5 2B on RTX 4070 Super report
RTX 50708-bit128K158 tok/s0.2 sPDF: MiniCPM5 2B on RTX 5070 report
RTX 4060 Ti 16 GB8-bit128K76 tok/s0.2 sPDF: MiniCPM5 2B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit128K113 tok/s0.2 sPDF: MiniCPM5 2B on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit128K90 tok/s0.2 sPDF: MiniCPM5 2B on Arc A770 16 GB report
Radeon RX 9070 XT8-bit128K120 tok/s0.1 sPDF: MiniCPM5 2B on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit128K158 tok/s0.1 sPDF: MiniCPM5 2B on RTX 4070 Ti Super report
RTX 4080 Super8-bit128K170 tok/s0.1 sPDF: MiniCPM5 2B on RTX 4080 Super report
RTX 5070 Ti8-bit128K197 tok/s0.1 sPDF: MiniCPM5 2B on RTX 5070 Ti report
RTX 50808-bit128K208 tok/s0.1 sPDF: MiniCPM5 2B on RTX 5080 report
Radeon RX 7900 XTX8-bit128K160 tok/s0.2 sPDF: MiniCPM5 2B on Radeon RX 7900 XTX report
RTX 30908-bit128K204 tok/s0.1 sPDF: MiniCPM5 2B on RTX 3090 report
RTX 40908-bit128K215 tok/s0.1 sPDF: MiniCPM5 2B on RTX 4090 report
Radeon AI PRO R97008-bit128K120 tok/s0.1 sPDF: MiniCPM5 2B on Radeon AI PRO R9700 report
RTX 50908-bit128K316 tok/s0.1 sPDF: MiniCPM5 2B on RTX 5090 report
2× RTX 30908-bit128K204 tok/s0.1 sPDF: MiniCPM5 2B on 2× RTX 3090 report
2× RTX 40908-bit128K215 tok/s0.1 sPDF: MiniCPM5 2B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit128K263 tok/s0.1 sPDF: MiniCPM5 2B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit128K208 tok/s0.1 sPDF: MiniCPM5 2B on RTX 6000 Ada report
L40S8-bit128K192 tok/s0.1 sPDF: MiniCPM5 2B on L40S report
2× RTX 50908-bit128K316 tok/s0.1 sPDF: MiniCPM5 2B on 2× RTX 5090 report
4× RTX 30908-bit128K204 tok/s0.1 sPDF: MiniCPM5 2B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit128K316 tok/s0.1 sPDF: MiniCPM5 2B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit128K36 tok/s3.0 sPDF: MiniCPM5 2B on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit128K36 tok/s3.0 sPDF: MiniCPM5 2B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit128K36 tok/s3.0 sPDF: MiniCPM5 2B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit128K71 tok/s1.5 sPDF: MiniCPM5 2B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit128K71 tok/s1.5 sPDF: MiniCPM5 2B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit128K71 tok/s1.5 sPDF: MiniCPM5 2B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit128K109 tok/s0.7 sPDF: MiniCPM5 2B on M4 Max 128 GB report
M5 Max 128 GB8-bit128K117 tok/s0.2 sPDF: MiniCPM5 2B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit128K127 tok/s0.4 sPDF: MiniCPM5 2B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit128K127 tok/s0.4 sPDF: MiniCPM5 2B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit128K127 tok/s0.4 sPDF: MiniCPM5 2B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit128K156 tok/s0.1 sPDF: MiniCPM5 2B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit128K57 tok/s0.4 sPDF: MiniCPM5 2B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit128K57 tok/s0.4 sPDF: MiniCPM5 2B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit128K79 tok/s0.1 sPDF: MiniCPM5 2B on RTX Spark report
DGX Spark8-bit128K73 tok/s0.1 sPDF: MiniCPM5 2B on DGX Spark report
Dual DGX Spark8-bit128K73 tok/s0.1 sPDF: MiniCPM5 2B on Dual DGX Spark report
DGX Station GB3008-bit128K574 tok/s0.1 sPDF: MiniCPM5 2B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit128K316 tok/s0.1 sPDF: MiniCPM5 2B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit128K316 tok/s0.1 sPDF: MiniCPM5 2B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit128K316 tok/s0.1 sPDF: MiniCPM5 2B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit128K341 tok/s0.1 sPDF: MiniCPM5 2B on A100 80 GB SXM report
H100 SXM8-bit128K439 tok/s0.1 sPDF: MiniCPM5 2B on H100 SXM report
4× H200 NVL8-bit128K507 tok/s0.1 sPDF: MiniCPM5 2B on 4× H200 NVL report
8× H200 NVL8-bit128K507 tok/s0.1 sPDF: MiniCPM5 2B on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

llama.cpp
llama-server -hf Abiray/MiniCPM5-2B-heretic-abliterated-GGUF
vLLM
vllm serve openbmb/MiniCPM5-2B
MLX (Mac)
mlx_lm.chat --model mlx-community/MiniCPM5-2B-mlx-4Bit

Questions

How much VRAM does MiniCPM5 2B need?

MiniCPM5 2B needs 1.4 GB for its weights at 4-bit, 2.5 GB with an 8K context and 3.5 GB with 32K. At 8-bit, 3.6 GB.

What is the minimum VRAM for MiniCPM5 2B?

About 2.2 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 4060.

How big is MiniCPM5 2B?

2.5B parameters: 4.7 GB of weights at 16-bit, 1.4 GB at 4-bit.

Can MiniCPM5 2B run on an RTX 4090?

Yes. At 8-bit it holds 128K of context and writes about 215 tokens a second.

More models