Skip to content

Qwen3 4B DFlash B16 VRAM requirements

How much VRAM Qwen3 4B DFlash B16 needs
4-bit weights: 311 MBWith 8K context: 1.3 GBWith 32K context: 1.7 GB8-bit, 8K context: 1.5 GB
Parameters
537M
Longest context
40K
Released
Jan 2026
License
MIT
Strongest at
Not rated

How much VRAM Qwen3 4B DFlash B16 needs

Memory Qwen3 4B DFlash B16 needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 40K
16-bit1.0 GB2.0 GB2.4 GB2.6 GB
8-bit545 MB1.5 GB2.0 GB2.1 GB
6-bit420 MB1.4 GB1.9 GB2.0 GB
5-bit365 MB1.3 GB1.8 GB2.0 GB
4-bit311 MB1.3 GB1.7 GB1.9 GB
3-bit250 MB1.2 GB1.7 GB1.8 GB
2-bit192 MB1.2 GB1.6 GB1.8 GB
1.6-bit106 MB1.1 GB1.5 GB1.7 GB

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

GPUs and Macs that run Qwen3 4B DFlash B16

The best way Qwen3 4B DFlash B16 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit40K318 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 4060 report
RTX 50608-bit40K507 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 5060 report
RTX 30708-bit40K507 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 3070 report
RTX 3080 10 GB8-bit40K817 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 3080 10 GB report
RTX 3060 12 GB8-bit40K414 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 3060 12 GB report
Arc B5808-bit40K357 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on Arc B580 report
RTX 40708-bit40K565 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 4070 report
RTX 4070 Super8-bit40K565 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 4070 Super report
RTX 50708-bit40K733 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 5070 report
RTX 4060 Ti 16 GB8-bit40K335 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit40K507 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit40K427 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on Arc A770 16 GB report
Radeon RX 9070 XT8-bit40K576 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit40K733 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 4070 Ti Super report
RTX 4080 Super8-bit40K795 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 4080 Super report
RTX 5070 Ti8-bit40K943 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 5070 Ti report
RTX 50808-bit40K1,000 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 5080 report
Radeon RX 7900 XTX8-bit40K806 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on Radeon RX 7900 XTX report
RTX 30908-bit40K979 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 3090 report
RTX 40908-bit40K1,042 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 4090 report
Radeon AI PRO R97008-bit40K576 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on Radeon AI PRO R9700 report
RTX 50908-bit40K1,652 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 5090 report
2× RTX 30908-bit40K979 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on 2× RTX 3090 report
2× RTX 40908-bit40K1,042 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit40K1,321 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit40K1,000 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX 6000 Ada report
L40S8-bit40K914 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on L40S report
2× RTX 50908-bit40K1,652 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on 2× RTX 5090 report
4× RTX 30908-bit40K979 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit40K1,652 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit40K161 tok/s0.4 sPDF: Qwen3 4B DFlash B16 on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit40K161 tok/s0.4 sPDF: Qwen3 4B DFlash B16 on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit40K161 tok/s0.4 sPDF: Qwen3 4B DFlash B16 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit40K339 tok/s0.2 sPDF: Qwen3 4B DFlash B16 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit40K339 tok/s0.2 sPDF: Qwen3 4B DFlash B16 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit40K339 tok/s0.2 sPDF: Qwen3 4B DFlash B16 on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit40K557 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on M4 Max 128 GB report
M5 Max 128 GB8-bit40K610 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on M5 Max 128 GB report
M3 Ultra 96 GB8-bit40K679 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit40K679 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit40K679 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit40K883 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit40K252 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit40K252 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit40K349 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on RTX Spark report
DGX Spark8-bit40K319 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on DGX Spark report
Dual DGX Spark8-bit40K319 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on Dual DGX Spark report
DGX Station GB3008-bit40K3,775 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit40K1,652 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit40K1,652 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit40K1,652 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit40K1,818 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on A100 80 GB SXM report
H100 SXM8-bit40K2,541 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on H100 SXM report
4× H200 NVL8-bit40K3,126 tok/s0.1 sPDF: Qwen3 4B DFlash B16 on 4× H200 NVL report
8× H200 NVL8-bit40K3,126 tok/s0.1 sPDF: Qwen3 4B DFlash B16 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 AtomicChat/Qwen3-4B-DFlash-GGUF
vLLM
vllm serve z-lab/Qwen3-4B-DFlash-b16

Questions

How much VRAM does Qwen3 4B DFlash B16 need?

Qwen3 4B DFlash B16 needs 311 MB for its weights at 4-bit, 1.3 GB with an 8K context and 1.7 GB with 32K. At 8-bit, 1.5 GB.

What is the minimum VRAM for Qwen3 4B DFlash B16?

About 1.2 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 Qwen3 4B DFlash B16?

537M parameters: 1.0 GB of weights at 16-bit, 311 MB at 4-bit.

Can Qwen3 4B DFlash B16 run on an RTX 4090?

Yes. At 8-bit it holds 40K of context and writes about 1,042 tokens a second.

More models