Skip to content

MiniCPM V 4.6 Thinking VRAM requirements

How much VRAM MiniCPM V 4.6 Thinking needs
4-bit weights: 752 MBWith 8K context: 1.8 GBWith 32K context: 2.1 GB8-bit, 8K context: 2.4 GB
Parameters
1.3B
Longest context
256K
Released
May 2026
License
Apache 2.0
Strongest at
Text and images

How much VRAM MiniCPM V 4.6 Thinking needs

Memory MiniCPM V 4.6 Thinking needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit2.4 GB3.5 GB3.8 GB4.9 GB
8-bit1.3 GB2.4 GB2.6 GB3.8 GB
6-bit1.0 GB2.1 GB2.3 GB3.5 GB
5-bit882 MB1.9 GB2.2 GB3.3 GB
4-bit752 MB1.8 GB2.1 GB3.2 GB
3-bit606 MB1.7 GB1.9 GB3.1 GB
2-bit465 MB1.5 GB1.8 GB2.9 GB
1.6-bit256 MB1.3 GB1.6 GB2.7 GB

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

GPUs and Macs that run MiniCPM V 4.6 Thinking

The best way MiniCPM V 4.6 Thinking runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit256K124 tok/s0.2 sPDF: MiniCPM V 4.6 Thinking on RTX 4060 report
RTX 50608-bit256K193 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 5060 report
RTX 30708-bit256K193 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 3070 report
RTX 3080 10 GB8-bit256K298 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 3080 10 GB report
RTX 3060 12 GB8-bit256K159 tok/s0.2 sPDF: MiniCPM V 4.6 Thinking on RTX 3060 12 GB report
Arc B5808-bit256K132 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on Arc B580 report
RTX 40708-bit256K213 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 4070 report
RTX 4070 Super8-bit256K213 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 4070 Super report
RTX 50708-bit256K270 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 5070 report
RTX 4060 Ti 16 GB8-bit256K130 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit256K193 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit256K155 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on Arc A770 16 GB report
Radeon RX 9070 XT8-bit256K206 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit256K270 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 4070 Ti Super report
RTX 4080 Super8-bit256K291 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 4080 Super report
RTX 5070 Ti8-bit256K339 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 5070 Ti report
RTX 50808-bit256K356 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 5080 report
Radeon RX 7900 XTX8-bit256K275 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on Radeon RX 7900 XTX report
RTX 30908-bit256K350 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 3090 report
RTX 40908-bit256K370 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 4090 report
Radeon AI PRO R97008-bit256K206 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on Radeon AI PRO R9700 report
RTX 50908-bit256K544 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 5090 report
2× RTX 30908-bit256K350 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on 2× RTX 3090 report
2× RTX 40908-bit256K370 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit256K453 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit256K356 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX 6000 Ada report
L40S8-bit256K329 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on L40S report
2× RTX 50908-bit256K544 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on 2× RTX 5090 report
4× RTX 30908-bit256K350 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit256K544 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit256K62 tok/s1.5 sPDF: MiniCPM V 4.6 Thinking on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit256K62 tok/s1.5 sPDF: MiniCPM V 4.6 Thinking on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit256K62 tok/s1.5 sPDF: MiniCPM V 4.6 Thinking on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit256K122 tok/s0.8 sPDF: MiniCPM V 4.6 Thinking on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit256K122 tok/s0.8 sPDF: MiniCPM V 4.6 Thinking on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit256K122 tok/s0.8 sPDF: MiniCPM V 4.6 Thinking on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit256K187 tok/s0.4 sPDF: MiniCPM V 4.6 Thinking on M4 Max 128 GB report
M5 Max 128 GB8-bit256K201 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on M5 Max 128 GB report
M3 Ultra 96 GB8-bit256K220 tok/s0.2 sPDF: MiniCPM V 4.6 Thinking on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit256K220 tok/s0.2 sPDF: MiniCPM V 4.6 Thinking on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit256K220 tok/s0.2 sPDF: MiniCPM V 4.6 Thinking on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit256K269 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit256K97 tok/s0.2 sPDF: MiniCPM V 4.6 Thinking on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit256K97 tok/s0.2 sPDF: MiniCPM V 4.6 Thinking on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit256K135 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on RTX Spark report
DGX Spark8-bit256K124 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on DGX Spark report
Dual DGX Spark8-bit256K124 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on Dual DGX Spark report
DGX Station GB3008-bit256K998 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit256K544 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit256K544 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit256K544 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit256K588 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on A100 80 GB SXM report
H100 SXM8-bit256K759 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on H100 SXM report
4× H200 NVL8-bit256K879 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking on 4× H200 NVL report
8× H200 NVL8-bit256K879 tok/s0.1 sPDF: MiniCPM V 4.6 Thinking 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 openbmb/MiniCPM-V-4.6-Thinking-gguf
vLLM
vllm serve openbmb/MiniCPM-V-4.6-Thinking

Questions

How much VRAM does MiniCPM V 4.6 Thinking need?

MiniCPM V 4.6 Thinking needs 752 MB for its weights at 4-bit, 1.8 GB with an 8K context and 2.1 GB with 32K. At 8-bit, 2.4 GB.

What is the minimum VRAM for MiniCPM V 4.6 Thinking?

About 1.7 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 MiniCPM V 4.6 Thinking?

1.3B parameters: 2.4 GB of weights at 16-bit, 752 MB at 4-bit.

Can MiniCPM V 4.6 Thinking run on an RTX 4090?

Yes. At 8-bit it holds 256K of context and writes about 370 tokens a second.

More models