Skip to content

Qwen2.5 32B Instruct VRAM requirements

How much VRAM Qwen2.5 32B Instruct needs
4-bit weights: 18.5 GBWith 8K context: 21.4 GBWith 32K context: 27.4 GB8-bit, 8K context: 35.3 GB
Parameters
32.8B
Longest context
32K
Released
Sep 2024
License
Apache 2.0
Strongest at
Not rated

How much VRAM Qwen2.5 32B Instruct needs

Memory Qwen2.5 32B Instruct needs at each precision and context length
PrecisionWeightsWith 8KWith 32K
16-bit61.0 GB63.9 GB69.9 GB
8-bit32.4 GB35.3 GB41.3 GB
6-bit25.0 GB27.9 GB33.9 GB
5-bit21.7 GB24.6 GB30.6 GB
4-bit18.5 GB21.4 GB27.4 GB
3-bit14.9 GB17.8 GB23.8 GB
2-bit11.4 GB14.3 GB20.3 GB
1.6-bit6.3 GB9.2 GB15.2 GB

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

GPUs and Macs that run Qwen2.5 32B Instruct

The best way Qwen2.5 32B Instruct runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Qwen2.5 32B Instruct on RTX 4060 report
RTX 5060Too bigPDF: Qwen2.5 32B Instruct on RTX 5060 report
RTX 3070Too bigPDF: Qwen2.5 32B Instruct on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Qwen2.5 32B Instruct on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: Qwen2.5 32B Instruct on RTX 3060 12 GB report
Arc B580Too bigPDF: Qwen2.5 32B Instruct on Arc B580 report
RTX 4070Too bigPDF: Qwen2.5 32B Instruct on RTX 4070 report
RTX 4070 SuperToo bigPDF: Qwen2.5 32B Instruct on RTX 4070 Super report
RTX 5070Too bigPDF: Qwen2.5 32B Instruct on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: Qwen2.5 32B Instruct on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: Qwen2.5 32B Instruct on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: Qwen2.5 32B Instruct on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: Qwen2.5 32B Instruct on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: Qwen2.5 32B Instruct on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: Qwen2.5 32B Instruct on RTX 4080 Super report
RTX 5070 TiToo bigPDF: Qwen2.5 32B Instruct on RTX 5070 Ti report
RTX 5080Too bigPDF: Qwen2.5 32B Instruct on RTX 5080 report
Radeon RX 7900 XTX3-bit24K33 tok/s3.0 sPDF: Qwen2.5 32B Instruct on Radeon RX 7900 XTX report
RTX 30903-bit24K39 tok/s1.9 sPDF: Qwen2.5 32B Instruct on RTX 3090 report
RTX 40903-bit24K42 tok/s0.8 sPDF: Qwen2.5 32B Instruct on RTX 4090 report
Radeon AI PRO R97006-bit16K14 tok/s2.0 sPDF: Qwen2.5 32B Instruct on Radeon AI PRO R9700 report
RTX 50906-bit16K44 tok/s0.7 sPDF: Qwen2.5 32B Instruct on RTX 5090 report
2× RTX 30908-bit32K19 tok/s1.1 sPDF: Qwen2.5 32B Instruct on 2× RTX 3090 report
2× RTX 40908-bit32K20 tok/s0.5 sPDF: Qwen2.5 32B Instruct on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit32K26 tok/s1.0 sPDF: Qwen2.5 32B Instruct on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit32K19 tok/s0.7 sPDF: Qwen2.5 32B Instruct on RTX 6000 Ada report
L40S8-bit32K17 tok/s1.2 sPDF: Qwen2.5 32B Instruct on L40S report
2× RTX 50908-bit32K34 tok/s0.4 sPDF: Qwen2.5 32B Instruct on 2× RTX 5090 report
4× RTX 30908-bit32K19 tok/s0.6 sPDF: Qwen2.5 32B Instruct on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit32K34 tok/s0.5 sPDF: Qwen2.5 32B Instruct on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Qwen2.5 32B Instruct on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: Qwen2.5 32B Instruct on Mac mini M4 24 GB report
Mac mini M4 32 GB3-bit16K6.1 tok/s44 sPDF: Qwen2.5 32B Instruct on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: Qwen2.5 32B Instruct on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB6-bit32K8.2 tok/s22 sPDF: Qwen2.5 32B Instruct on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit32K6.4 tok/s22 sPDF: Qwen2.5 32B Instruct on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit32K11 tok/s11 sPDF: Qwen2.5 32B Instruct on M4 Max 128 GB report
M5 Max 128 GB8-bit32K13 tok/s3.2 sPDF: Qwen2.5 32B Instruct on M5 Max 128 GB report
M3 Ultra 96 GB8-bit32K15 tok/s6.1 sPDF: Qwen2.5 32B Instruct on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit32K15 tok/s6.1 sPDF: Qwen2.5 32B Instruct on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit32K15 tok/s6.1 sPDF: Qwen2.5 32B Instruct on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit32K21 tok/s1.9 sPDF: Qwen2.5 32B Instruct on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit32K4.5 tok/s6.3 sPDF: Qwen2.5 32B Instruct on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit32K4.5 tok/s6.3 sPDF: Qwen2.5 32B Instruct on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit32K6.1 tok/s2.2 sPDF: Qwen2.5 32B Instruct on RTX Spark report
DGX Spark8-bit32K5.6 tok/s2.2 sPDF: Qwen2.5 32B Instruct on DGX Spark report
Dual DGX Spark8-bit32K5.6 tok/s1.3 sPDF: Qwen2.5 32B Instruct on Dual DGX Spark report
DGX Station GB3008-bit32K114 tok/s0.2 sPDF: Qwen2.5 32B Instruct on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit32K34 tok/s0.3 sPDF: Qwen2.5 32B Instruct on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit32K34 tok/s0.2 sPDF: Qwen2.5 32B Instruct on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit32K34 tok/s0.1 sPDF: Qwen2.5 32B Instruct on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit32K39 tok/s1.4 sPDF: Qwen2.5 32B Instruct on A100 80 GB SXM report
H100 SXM8-bit32K61 tok/s0.4 sPDF: Qwen2.5 32B Instruct on H100 SXM report
4× H200 NVL8-bit32K83 tok/s0.2 sPDF: Qwen2.5 32B Instruct on 4× H200 NVL report
8× H200 NVL8-bit32K83 tok/s0.1 sPDF: Qwen2.5 32B Instruct 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 bartowski/Qwen2.5-32B-Instruct-GGUF
vLLM
vllm serve Qwen/Qwen2.5-32B-Instruct

Questions

How much VRAM does Qwen2.5 32B Instruct need?

Qwen2.5 32B Instruct needs 18.5 GB for its weights at 4-bit, 21.4 GB with an 8K context and 27.4 GB with 32K. At 8-bit, 35.3 GB.

What is the minimum VRAM for Qwen2.5 32B Instruct?

About 17.8 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 Radeon RX 7900 XTX.

How big is Qwen2.5 32B Instruct?

32.8B parameters: 61.0 GB of weights at 16-bit, 18.5 GB at 4-bit.

Can Qwen2.5 32B Instruct run on an RTX 4090?

Yes. At 3-bit it holds 24K of context and writes about 42 tokens a second.

More models