Skip to content

Qwen3 4B Instruct 2507 VRAM requirements

How much VRAM Qwen3 4B Instruct 2507 needs
4-bit weights: 2.3 GBWith 8K context: 4.2 GBWith 32K context: 7.6 GB8-bit, 8K context: 5.9 GB
Parameters
4B
Longest context
256K
Released
Aug 2025
License
Apache 2.0
Strongest at
Not rated

How much VRAM Qwen3 4B Instruct 2507 needs

Memory Qwen3 4B Instruct 2507 needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit7.5 GB9.4 GB12.8 GB26.3 GB
8-bit4.0 GB5.9 GB9.3 GB22.8 GB
6-bit3.1 GB5.0 GB8.4 GB21.9 GB
5-bit2.7 GB4.6 GB8.0 GB21.5 GB
4-bit2.3 GB4.2 GB7.6 GB21.1 GB
3-bit1.8 GB3.8 GB7.1 GB20.6 GB
2-bit1.4 GB3.3 GB6.7 GB20.2 GB
1.6-bit791 MB2.7 GB6.1 GB19.6 GB

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

GPUs and Macs that run Qwen3 4B Instruct 2507

The best way Qwen3 4B Instruct 2507 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit16K43 tok/s0.5 sPDF: Qwen3 4B Instruct 2507 on RTX 4060 report
RTX 50608-bit16K68 tok/s0.4 sPDF: Qwen3 4B Instruct 2507 on RTX 5060 report
RTX 30708-bit16K68 tok/s0.4 sPDF: Qwen3 4B Instruct 2507 on RTX 3070 report
RTX 3080 10 GB8-bit32K110 tok/s0.3 sPDF: Qwen3 4B Instruct 2507 on RTX 3080 10 GB report
RTX 3060 12 GB8-bit44K55 tok/s0.6 sPDF: Qwen3 4B Instruct 2507 on RTX 3060 12 GB report
Arc B5808-bit44K48 tok/s0.4 sPDF: Qwen3 4B Instruct 2507 on Arc B580 report
RTX 40708-bit44K76 tok/s0.3 sPDF: Qwen3 4B Instruct 2507 on RTX 4070 report
RTX 4070 Super8-bit44K76 tok/s0.2 sPDF: Qwen3 4B Instruct 2507 on RTX 4070 Super report
RTX 50708-bit44K98 tok/s0.3 sPDF: Qwen3 4B Instruct 2507 on RTX 5070 report
RTX 4060 Ti 16 GB8-bit72K45 tok/s0.4 sPDF: Qwen3 4B Instruct 2507 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit72K68 tok/s0.3 sPDF: Qwen3 4B Instruct 2507 on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit72K57 tok/s0.4 sPDF: Qwen3 4B Instruct 2507 on Arc A770 16 GB report
Radeon RX 9070 XT8-bit72K77 tok/s0.2 sPDF: Qwen3 4B Instruct 2507 on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit72K98 tok/s0.2 sPDF: Qwen3 4B Instruct 2507 on RTX 4070 Ti Super report
RTX 4080 Super8-bit72K107 tok/s0.2 sPDF: Qwen3 4B Instruct 2507 on RTX 4080 Super report
RTX 5070 Ti8-bit72K127 tok/s0.2 sPDF: Qwen3 4B Instruct 2507 on RTX 5070 Ti report
RTX 50808-bit72K134 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on RTX 5080 report
Radeon RX 7900 XTX8-bit128K108 tok/s0.4 sPDF: Qwen3 4B Instruct 2507 on Radeon RX 7900 XTX report
RTX 30908-bit128K132 tok/s0.2 sPDF: Qwen3 4B Instruct 2507 on RTX 3090 report
RTX 40908-bit128K140 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on RTX 4090 report
Radeon AI PRO R97008-bit176K77 tok/s0.2 sPDF: Qwen3 4B Instruct 2507 on Radeon AI PRO R9700 report
RTX 50908-bit176K223 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on RTX 5090 report
2× RTX 30908-bit256K132 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on 2× RTX 3090 report
2× RTX 40908-bit256K140 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit256K178 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit256K134 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on RTX 6000 Ada report
L40S8-bit256K123 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on L40S report
2× RTX 50908-bit256K223 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on 2× RTX 5090 report
4× RTX 30908-bit256K132 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit256K223 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit36K22 tok/s5.2 sPDF: Qwen3 4B Instruct 2507 on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit72K22 tok/s5.2 sPDF: Qwen3 4B Instruct 2507 on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit104K22 tok/s5.2 sPDF: Qwen3 4B Instruct 2507 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit72K46 tok/s2.6 sPDF: Qwen3 4B Instruct 2507 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit208K46 tok/s2.6 sPDF: Qwen3 4B Instruct 2507 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit256K46 tok/s2.6 sPDF: Qwen3 4B Instruct 2507 on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit256K75 tok/s1.3 sPDF: Qwen3 4B Instruct 2507 on M4 Max 128 GB report
M5 Max 128 GB8-bit256K82 tok/s0.4 sPDF: Qwen3 4B Instruct 2507 on M5 Max 128 GB report
M3 Ultra 96 GB8-bit256K92 tok/s0.7 sPDF: Qwen3 4B Instruct 2507 on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit256K92 tok/s0.7 sPDF: Qwen3 4B Instruct 2507 on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit256K92 tok/s0.7 sPDF: Qwen3 4B Instruct 2507 on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit256K120 tok/s0.2 sPDF: Qwen3 4B Instruct 2507 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit256K34 tok/s0.8 sPDF: Qwen3 4B Instruct 2507 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit256K34 tok/s0.8 sPDF: Qwen3 4B Instruct 2507 on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit256K47 tok/s0.3 sPDF: Qwen3 4B Instruct 2507 on RTX Spark report
DGX Spark8-bit256K43 tok/s0.3 sPDF: Qwen3 4B Instruct 2507 on DGX Spark report
Dual DGX Spark8-bit256K43 tok/s0.2 sPDF: Qwen3 4B Instruct 2507 on Dual DGX Spark report
DGX Station GB3008-bit256K515 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit256K223 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit256K223 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit256K223 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit256K245 tok/s0.2 sPDF: Qwen3 4B Instruct 2507 on A100 80 GB SXM report
H100 SXM8-bit256K345 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on H100 SXM report
4× H200 NVL8-bit256K425 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 on 4× H200 NVL report
8× H200 NVL8-bit256K425 tok/s0.1 sPDF: Qwen3 4B Instruct 2507 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/Qwen3-4B-Instruct-2507-GGUF
vLLM
vllm serve Qwen/Qwen3-4B-Instruct-2507
MLX (Mac)
mlx_lm.chat --model mlx-community/Qwen3-4B-Instruct-2507-4bit

Questions

How much VRAM does Qwen3 4B Instruct 2507 need?

Qwen3 4B Instruct 2507 needs 2.3 GB for its weights at 4-bit, 4.2 GB with an 8K context and 7.6 GB with 32K. At 8-bit, 5.9 GB.

What is the minimum VRAM for Qwen3 4B Instruct 2507?

About 3.8 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 RTX 4060.

How big is Qwen3 4B Instruct 2507?

4B parameters: 7.5 GB of weights at 16-bit, 2.3 GB at 4-bit.

Can Qwen3 4B Instruct 2507 run on an RTX 4090?

Yes. At 8-bit it holds 128K of context and writes about 140 tokens a second.

More models