Skip to content

Qwen1.5 MOE A2.7B VRAM requirements

How much VRAM Qwen1.5 MOE A2.7B needs
4-bit weights: 8.1 GBWith 8K context: 10.4 GB8-bit, 8K context: 16.5 GB
Parameters
14.3B
Active per token
2.7B
Longest context
8K
Released
Feb 2024
License
Tongyi Qianwen
Strongest at
Not rated

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

How much VRAM Qwen1.5 MOE A2.7B needs

Memory Qwen1.5 MOE A2.7B needs at each precision and context length
PrecisionWeightsWith 8K
16-bit26.7 GB29.0 GB
8-bit14.2 GB16.5 GB
6-bit10.9 GB13.2 GB
5-bit9.5 GB11.8 GB
4-bit8.1 GB10.4 GB
3-bit6.5 GB8.8 GB
2-bit5.0 GB7.3 GB
1.6-bit2.7 GB5.1 GB

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

GPUs and Macs that run Qwen1.5 MOE A2.7B

The best way Qwen1.5 MOE A2.7B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Qwen1.5 MOE A2.7B on RTX 4060 report
RTX 5060Too bigPDF: Qwen1.5 MOE A2.7B on RTX 5060 report
RTX 3070Too bigPDF: Qwen1.5 MOE A2.7B on RTX 3070 report
RTX 3080 10 GB3-bit8K255 tok/s0.5 sPDF: Qwen1.5 MOE A2.7B on RTX 3080 10 GB report
RTX 3060 12 GB4-bit8K126 tok/s1.2 sPDF: Qwen1.5 MOE A2.7B on RTX 3060 12 GB report
Arc B5804-bit8K104 tok/s0.8 sPDF: Qwen1.5 MOE A2.7B on Arc B580 report
RTX 40704-bit8K164 tok/s0.5 sPDF: Qwen1.5 MOE A2.7B on RTX 4070 report
RTX 4070 Super4-bit8K164 tok/s0.4 sPDF: Qwen1.5 MOE A2.7B on RTX 4070 Super report
RTX 50704-bit8K203 tok/s0.5 sPDF: Qwen1.5 MOE A2.7B on RTX 5070 report
RTX 4060 Ti 16 GB6-bit8K81 tok/s0.7 sPDF: Qwen1.5 MOE A2.7B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB6-bit8K118 tok/s0.7 sPDF: Qwen1.5 MOE A2.7B on RTX 5060 Ti 16 GB report
Arc A770 16 GB6-bit8K97 tok/s0.8 sPDF: Qwen1.5 MOE A2.7B on Arc A770 16 GB report
Radeon RX 9070 XT6-bit8K126 tok/s0.4 sPDF: Qwen1.5 MOE A2.7B on Radeon RX 9070 XT report
RTX 4070 Ti Super6-bit8K163 tok/s0.4 sPDF: Qwen1.5 MOE A2.7B on RTX 4070 Ti Super report
RTX 4080 Super6-bit8K174 tok/s0.3 sPDF: Qwen1.5 MOE A2.7B on RTX 4080 Super report
RTX 5070 Ti6-bit8K201 tok/s0.4 sPDF: Qwen1.5 MOE A2.7B on RTX 5070 Ti report
RTX 50806-bit8K211 tok/s0.3 sPDF: Qwen1.5 MOE A2.7B on RTX 5080 report
Radeon RX 7900 XTX8-bit8K140 tok/s0.7 sPDF: Qwen1.5 MOE A2.7B on Radeon RX 7900 XTX report
RTX 30908-bit8K172 tok/s0.4 sPDF: Qwen1.5 MOE A2.7B on RTX 3090 report
RTX 40908-bit8K182 tok/s0.2 sPDF: Qwen1.5 MOE A2.7B on RTX 4090 report
Radeon AI PRO R97008-bit8K104 tok/s0.4 sPDF: Qwen1.5 MOE A2.7B on Radeon AI PRO R9700 report
RTX 50908-bit8K268 tok/s0.1 sPDF: Qwen1.5 MOE A2.7B on RTX 5090 report
2× RTX 30908-bit8K172 tok/s0.3 sPDF: Qwen1.5 MOE A2.7B on 2× RTX 3090 report
2× RTX 40908-bit8K182 tok/s0.1 sPDF: Qwen1.5 MOE A2.7B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit8K223 tok/s0.2 sPDF: Qwen1.5 MOE A2.7B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit8K175 tok/s0.2 sPDF: Qwen1.5 MOE A2.7B on RTX 6000 Ada report
L40S8-bit8K162 tok/s0.3 sPDF: Qwen1.5 MOE A2.7B on L40S report
2× RTX 50908-bit8K268 tok/s0.1 sPDF: Qwen1.5 MOE A2.7B on 2× RTX 5090 report
4× RTX 30908-bit8K172 tok/s0.1 sPDF: Qwen1.5 MOE A2.7B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit8K268 tok/s0.1 sPDF: Qwen1.5 MOE A2.7B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB3-bit8K60 tok/s9.9 sPDF: Qwen1.5 MOE A2.7B on Mac mini M4 16 GB report
Mac mini M4 24 GB6-bit8K39 tok/s9.9 sPDF: Qwen1.5 MOE A2.7B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit8K31 tok/s9.9 sPDF: Qwen1.5 MOE A2.7B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB6-bit8K75 tok/s5.0 sPDF: Qwen1.5 MOE A2.7B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit8K62 tok/s5.0 sPDF: Qwen1.5 MOE A2.7B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit8K62 tok/s5.0 sPDF: Qwen1.5 MOE A2.7B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit8K96 tok/s2.5 sPDF: Qwen1.5 MOE A2.7B on M4 Max 128 GB report
M5 Max 128 GB8-bit8K104 tok/s0.7 sPDF: Qwen1.5 MOE A2.7B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit8K114 tok/s1.4 sPDF: Qwen1.5 MOE A2.7B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit8K114 tok/s1.4 sPDF: Qwen1.5 MOE A2.7B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit8K114 tok/s1.4 sPDF: Qwen1.5 MOE A2.7B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit8K142 tok/s0.4 sPDF: Qwen1.5 MOE A2.7B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit8K48 tok/s1.4 sPDF: Qwen1.5 MOE A2.7B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit8K48 tok/s1.4 sPDF: Qwen1.5 MOE A2.7B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit8K66 tok/s0.5 sPDF: Qwen1.5 MOE A2.7B on RTX Spark report
DGX Spark8-bit8K61 tok/s0.5 sPDF: Qwen1.5 MOE A2.7B on DGX Spark report
Dual DGX Spark8-bit8K61 tok/s0.3 sPDF: Qwen1.5 MOE A2.7B on Dual DGX Spark report
DGX Station GB3008-bit8K496 tok/s0.1 sPDF: Qwen1.5 MOE A2.7B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit8K268 tok/s0.1 sPDF: Qwen1.5 MOE A2.7B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit8K268 tok/s0.1 sPDF: Qwen1.5 MOE A2.7B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit8K268 tok/s0.1 sPDF: Qwen1.5 MOE A2.7B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit8K290 tok/s0.3 sPDF: Qwen1.5 MOE A2.7B on A100 80 GB SXM report
H100 SXM8-bit8K375 tok/s0.1 sPDF: Qwen1.5 MOE A2.7B on H100 SXM report
4× H200 NVL8-bit8K436 tok/s0.1 sPDF: Qwen1.5 MOE A2.7B on 4× H200 NVL report
8× H200 NVL8-bit8K436 tok/s0.1 sPDF: Qwen1.5 MOE A2.7B 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 QuantFactory/Qwen1.5-MoE-A2.7B-Wikihow-GGUF
vLLM
vllm serve Qwen/Qwen1.5-MoE-A2.7B

Questions

How much VRAM does Qwen1.5 MOE A2.7B need?

Qwen1.5 MOE A2.7B needs 8.1 GB for its weights at 4-bit, 10.4 GB with an 8K context and 10.4 GB with 8K. At 8-bit, 16.5 GB.

What is the minimum VRAM for Qwen1.5 MOE A2.7B?

About 8.8 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 3080 10 GB.

How big is Qwen1.5 MOE A2.7B?

14.3B parameters (2.7B active per token): 26.7 GB of weights at 16-bit, 8.1 GB at 4-bit.

Can Qwen1.5 MOE A2.7B run on an RTX 4090?

Yes. At 8-bit it holds 8K of context and writes about 182 tokens a second.

More models