Skip to content

Qwen3 VL 2B Thinking VRAM requirements

How much VRAM Qwen3 VL 2B Thinking needs
4-bit weights: 1.2 GBWith 8K context: 2.9 GBWith 32K context: 5.5 GB8-bit, 8K context: 3.8 GB
Parameters
2.1B
Longest context
256K
Released
Oct 2025
License
Apache 2.0
Strongest at
Text and images

How much VRAM Qwen3 VL 2B Thinking needs

Memory Qwen3 VL 2B Thinking needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit4.0 GB5.6 GB8.3 GB18.8 GB
8-bit2.1 GB3.8 GB6.4 GB16.9 GB
6-bit1.6 GB3.3 GB5.9 GB16.4 GB
5-bit1.4 GB3.1 GB5.7 GB16.2 GB
4-bit1.2 GB2.9 GB5.5 GB16.0 GB
3-bit1.0 GB2.6 GB5.3 GB15.8 GB
2-bit761 MB2.4 GB5.0 GB15.5 GB
1.6-bit418 MB2.1 GB4.7 GB15.2 GB

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

GPUs and Macs that run Qwen3 VL 2B Thinking

The best way Qwen3 VL 2B Thinking runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit40K79 tok/s0.3 sPDF: Qwen3 VL 2B Thinking on RTX 4060 report
RTX 50608-bit40K124 tok/s0.2 sPDF: Qwen3 VL 2B Thinking on RTX 5060 report
RTX 30708-bit40K124 tok/s0.2 sPDF: Qwen3 VL 2B Thinking on RTX 3070 report
RTX 3080 10 GB8-bit60K196 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX 3080 10 GB report
RTX 3060 12 GB8-bit72K102 tok/s0.3 sPDF: Qwen3 VL 2B Thinking on RTX 3060 12 GB report
Arc B5808-bit72K86 tok/s0.2 sPDF: Qwen3 VL 2B Thinking on Arc B580 report
RTX 40708-bit72K138 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX 4070 report
RTX 4070 Super8-bit72K138 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX 4070 Super report
RTX 50708-bit72K177 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX 5070 report
RTX 4060 Ti 16 GB8-bit112K83 tok/s0.2 sPDF: Qwen3 VL 2B Thinking on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit112K124 tok/s0.2 sPDF: Qwen3 VL 2B Thinking on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit112K103 tok/s0.2 sPDF: Qwen3 VL 2B Thinking on Arc A770 16 GB report
Radeon RX 9070 XT8-bit112K137 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit112K177 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX 4070 Ti Super report
RTX 4080 Super8-bit112K191 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX 4080 Super report
RTX 5070 Ti8-bit112K225 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX 5070 Ti report
RTX 50808-bit112K238 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX 5080 report
Radeon RX 7900 XTX8-bit176K188 tok/s0.2 sPDF: Qwen3 VL 2B Thinking on Radeon RX 7900 XTX report
RTX 30908-bit176K233 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX 3090 report
RTX 40908-bit176K247 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX 4090 report
Radeon AI PRO R97008-bit248K137 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on Radeon AI PRO R9700 report
RTX 50908-bit248K378 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX 5090 report
2× RTX 30908-bit256K233 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on 2× RTX 3090 report
2× RTX 40908-bit256K247 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit256K308 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit256K238 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX 6000 Ada report
L40S8-bit256K218 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on L40S report
2× RTX 50908-bit256K378 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on 2× RTX 5090 report
4× RTX 30908-bit256K233 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit256K378 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit64K40 tok/s2.6 sPDF: Qwen3 VL 2B Thinking on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit112K40 tok/s2.6 sPDF: Qwen3 VL 2B Thinking on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit160K40 tok/s2.6 sPDF: Qwen3 VL 2B Thinking on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit112K81 tok/s1.3 sPDF: Qwen3 VL 2B Thinking on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit256K81 tok/s1.3 sPDF: Qwen3 VL 2B Thinking on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit256K81 tok/s1.3 sPDF: Qwen3 VL 2B Thinking on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit256K129 tok/s0.7 sPDF: Qwen3 VL 2B Thinking on M4 Max 128 GB report
M5 Max 128 GB8-bit256K140 tok/s0.2 sPDF: Qwen3 VL 2B Thinking on M5 Max 128 GB report
M3 Ultra 96 GB8-bit256K154 tok/s0.4 sPDF: Qwen3 VL 2B Thinking on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit256K154 tok/s0.4 sPDF: Qwen3 VL 2B Thinking on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit256K154 tok/s0.4 sPDF: Qwen3 VL 2B Thinking on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit256K195 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit256K62 tok/s0.4 sPDF: Qwen3 VL 2B Thinking on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit256K62 tok/s0.4 sPDF: Qwen3 VL 2B Thinking on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit256K86 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on RTX Spark report
DGX Spark8-bit256K79 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on DGX Spark report
Dual DGX Spark8-bit256K79 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on Dual DGX Spark report
DGX Station GB3008-bit256K772 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit256K378 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit256K378 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit256K378 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit256K413 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on A100 80 GB SXM report
H100 SXM8-bit256K554 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on H100 SXM report
4× H200 NVL8-bit256K661 tok/s0.1 sPDF: Qwen3 VL 2B Thinking on 4× H200 NVL report
8× H200 NVL8-bit256K661 tok/s0.1 sPDF: Qwen3 VL 2B 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 unsloth/Qwen3-VL-2B-Thinking-GGUF
vLLM
vllm serve Qwen/Qwen3-VL-2B-Thinking

Questions

How much VRAM does Qwen3 VL 2B Thinking need?

Qwen3 VL 2B Thinking needs 1.2 GB for its weights at 4-bit, 2.9 GB with an 8K context and 5.5 GB with 32K. At 8-bit, 3.8 GB.

What is the minimum VRAM for Qwen3 VL 2B Thinking?

About 2.6 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 Qwen3 VL 2B Thinking?

2.1B parameters: 4.0 GB of weights at 16-bit, 1.2 GB at 4-bit.

Can Qwen3 VL 2B Thinking run on an RTX 4090?

Yes. At 8-bit it holds 176K of context and writes about 247 tokens a second.

More models