Skip to content

Qwen1.5 4B Chat VRAM requirements

How much VRAM Qwen1.5 4B Chat needs
4-bit weights: 2.2 GBWith 8K context: 6.2 GBWith 32K context: 15.5 GB8-bit, 8K context: 7.9 GB
Parameters
4B
Longest context
32K
Released
Jan 2024
License
Tongyi Qianwen Research
Strongest at
Not rated

Its licence (Tongyi Qianwen Research) is not a standard open-source one: check the terms before commercial use.

How much VRAM Qwen1.5 4B Chat needs

Memory Qwen1.5 4B Chat needs at each precision and context length
PrecisionWeightsWith 8KWith 32K
16-bit7.4 GB11.3 GB20.7 GB
8-bit3.9 GB7.9 GB17.2 GB
6-bit3.0 GB7.0 GB16.3 GB
5-bit2.6 GB6.6 GB15.9 GB
4-bit2.2 GB6.2 GB15.5 GB
3-bit1.8 GB5.7 GB15.1 GB
2-bit1.4 GB5.3 GB14.7 GB
1.6-bit777 MB4.7 GB14.1 GB

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

GPUs and Macs that run Qwen1.5 4B Chat

The best way Qwen1.5 4B Chat runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Qwen1.5 4B Chat on RTX 4060 report
RTX 5060Too bigPDF: Qwen1.5 4B Chat on RTX 5060 report
RTX 3070Too bigPDF: Qwen1.5 4B Chat on RTX 3070 report
RTX 3080 10 GB4-bit16K190 tok/s0.2 sPDF: Qwen1.5 4B Chat on RTX 3080 10 GB report
RTX 3060 12 GB8-bit16K62 tok/s0.6 sPDF: Qwen1.5 4B Chat on RTX 3060 12 GB report
Arc B5808-bit16K53 tok/s0.4 sPDF: Qwen1.5 4B Chat on Arc B580 report
RTX 40708-bit16K84 tok/s0.3 sPDF: Qwen1.5 4B Chat on RTX 4070 report
RTX 4070 Super8-bit16K84 tok/s0.2 sPDF: Qwen1.5 4B Chat on RTX 4070 Super report
RTX 50708-bit16K108 tok/s0.2 sPDF: Qwen1.5 4B Chat on RTX 5070 report
RTX 4060 Ti 16 GB8-bit24K50 tok/s0.3 sPDF: Qwen1.5 4B Chat on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit24K75 tok/s0.3 sPDF: Qwen1.5 4B Chat on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit24K63 tok/s0.4 sPDF: Qwen1.5 4B Chat on Arc A770 16 GB report
Radeon RX 9070 XT8-bit24K84 tok/s0.2 sPDF: Qwen1.5 4B Chat on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit24K108 tok/s0.2 sPDF: Qwen1.5 4B Chat on RTX 4070 Ti Super report
RTX 4080 Super8-bit24K117 tok/s0.1 sPDF: Qwen1.5 4B Chat on RTX 4080 Super report
RTX 5070 Ti8-bit24K138 tok/s0.2 sPDF: Qwen1.5 4B Chat on RTX 5070 Ti report
RTX 50808-bit24K146 tok/s0.1 sPDF: Qwen1.5 4B Chat on RTX 5080 report
Radeon RX 7900 XTX8-bit32K117 tok/s0.3 sPDF: Qwen1.5 4B Chat on Radeon RX 7900 XTX report
RTX 30908-bit32K143 tok/s0.2 sPDF: Qwen1.5 4B Chat on RTX 3090 report
RTX 40908-bit32K152 tok/s0.1 sPDF: Qwen1.5 4B Chat on RTX 4090 report
Radeon AI PRO R97008-bit32K84 tok/s0.2 sPDF: Qwen1.5 4B Chat on Radeon AI PRO R9700 report
RTX 50908-bit32K237 tok/s0.1 sPDF: Qwen1.5 4B Chat on RTX 5090 report
2× RTX 30908-bit32K143 tok/s0.1 sPDF: Qwen1.5 4B Chat on 2× RTX 3090 report
2× RTX 40908-bit32K152 tok/s0.1 sPDF: Qwen1.5 4B Chat on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit32K191 tok/s0.1 sPDF: Qwen1.5 4B Chat on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit32K146 tok/s0.1 sPDF: Qwen1.5 4B Chat on RTX 6000 Ada report
L40S8-bit32K134 tok/s0.1 sPDF: Qwen1.5 4B Chat on L40S report
2× RTX 50908-bit32K237 tok/s0.1 sPDF: Qwen1.5 4B Chat on 2× RTX 5090 report
4× RTX 30908-bit32K143 tok/s0.1 sPDF: Qwen1.5 4B Chat on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit32K237 tok/s0.1 sPDF: Qwen1.5 4B Chat on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB6-bit16K30 tok/s4.6 sPDF: Qwen1.5 4B Chat on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit24K24 tok/s4.6 sPDF: Qwen1.5 4B Chat on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit32K24 tok/s4.6 sPDF: Qwen1.5 4B Chat on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit24K50 tok/s2.3 sPDF: Qwen1.5 4B Chat on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit32K50 tok/s2.3 sPDF: Qwen1.5 4B Chat on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit32K50 tok/s2.3 sPDF: Qwen1.5 4B Chat on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit32K80 tok/s1.2 sPDF: Qwen1.5 4B Chat on M4 Max 128 GB report
M5 Max 128 GB8-bit32K88 tok/s0.3 sPDF: Qwen1.5 4B Chat on M5 Max 128 GB report
M3 Ultra 96 GB8-bit32K97 tok/s0.6 sPDF: Qwen1.5 4B Chat on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit32K97 tok/s0.6 sPDF: Qwen1.5 4B Chat on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit32K97 tok/s0.6 sPDF: Qwen1.5 4B Chat on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit32K124 tok/s0.2 sPDF: Qwen1.5 4B Chat on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit32K38 tok/s0.7 sPDF: Qwen1.5 4B Chat on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit32K38 tok/s0.7 sPDF: Qwen1.5 4B Chat on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit32K52 tok/s0.2 sPDF: Qwen1.5 4B Chat on RTX Spark report
DGX Spark8-bit32K48 tok/s0.2 sPDF: Qwen1.5 4B Chat on DGX Spark report
Dual DGX Spark8-bit32K48 tok/s0.1 sPDF: Qwen1.5 4B Chat on Dual DGX Spark report
DGX Station GB3008-bit32K510 tok/s0.1 sPDF: Qwen1.5 4B Chat on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit32K237 tok/s0.1 sPDF: Qwen1.5 4B Chat on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit32K237 tok/s0.1 sPDF: Qwen1.5 4B Chat on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit32K237 tok/s0.1 sPDF: Qwen1.5 4B Chat on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit32K260 tok/s0.1 sPDF: Qwen1.5 4B Chat on A100 80 GB SXM report
H100 SXM8-bit32K355 tok/s0.1 sPDF: Qwen1.5 4B Chat on H100 SXM report
4× H200 NVL8-bit32K430 tok/s0.1 sPDF: Qwen1.5 4B Chat on 4× H200 NVL report
8× H200 NVL8-bit32K430 tok/s0.1 sPDF: Qwen1.5 4B Chat 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 second-state/Qwen1.5-4B-Chat-GGUF
vLLM
vllm serve Qwen/Qwen1.5-4B-Chat

Questions

How much VRAM does Qwen1.5 4B Chat need?

Qwen1.5 4B Chat needs 2.2 GB for its weights at 4-bit, 6.2 GB with an 8K context and 15.5 GB with 32K. At 8-bit, 7.9 GB.

What is the minimum VRAM for Qwen1.5 4B Chat?

About 5.7 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 3080 10 GB.

How big is Qwen1.5 4B Chat?

4B parameters: 7.4 GB of weights at 16-bit, 2.2 GB at 4-bit.

Can Qwen1.5 4B Chat run on an RTX 4090?

Yes. At 8-bit it holds 32K of context and writes about 152 tokens a second.

More models