Skip to content

Qwen2 57B A14B Instruct VRAM requirements

How much VRAM Qwen2 57B A14B Instruct needs
4-bit weights: 32.4 GBWith 8K context: 33.7 GBWith 32K context: 35.0 GB8-bit, 8K context: 58.1 GB
Parameters
57.4B
Active per token
14B
Longest context
32K
Released
Jun 2024
License
Apache 2.0
Strongest at
Not rated

How much VRAM Qwen2 57B A14B Instruct needs

Memory Qwen2 57B A14B Instruct needs at each precision and context length
PrecisionWeightsWith 8KWith 32K
16-bit107 GB108 GB110 GB
8-bit56.8 GB58.1 GB59.4 GB
6-bit43.8 GB45.1 GB46.4 GB
5-bit38.0 GB39.3 GB40.6 GB
4-bit32.4 GB33.7 GB35.0 GB
3-bit26.1 GB27.4 GB28.7 GB
2-bit20.0 GB21.3 GB22.6 GB
1.6-bit11.0 GB12.3 GB13.6 GB

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

GPUs and Macs that run Qwen2 57B A14B Instruct

The best way Qwen2 57B A14B Instruct runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Qwen2 57B A14B Instruct on RTX 4060 report
RTX 5060Too bigPDF: Qwen2 57B A14B Instruct on RTX 5060 report
RTX 3070Too bigPDF: Qwen2 57B A14B Instruct on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Qwen2 57B A14B Instruct on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: Qwen2 57B A14B Instruct on RTX 3060 12 GB report
Arc B580Too bigPDF: Qwen2 57B A14B Instruct on Arc B580 report
RTX 4070Too bigPDF: Qwen2 57B A14B Instruct on RTX 4070 report
RTX 4070 SuperToo bigPDF: Qwen2 57B A14B Instruct on RTX 4070 Super report
RTX 5070Too bigPDF: Qwen2 57B A14B Instruct on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: Qwen2 57B A14B Instruct on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: Qwen2 57B A14B Instruct on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: Qwen2 57B A14B Instruct on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: Qwen2 57B A14B Instruct on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: Qwen2 57B A14B Instruct on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: Qwen2 57B A14B Instruct on RTX 4080 Super report
RTX 5070 TiToo bigPDF: Qwen2 57B A14B Instruct on RTX 5070 Ti report
RTX 5080Too bigPDF: Qwen2 57B A14B Instruct on RTX 5080 report
Radeon RX 7900 XTXToo bigPDF: Qwen2 57B A14B Instruct on Radeon RX 7900 XTX report
RTX 3090Too bigPDF: Qwen2 57B A14B Instruct on RTX 3090 report
RTX 4090Too bigPDF: Qwen2 57B A14B Instruct on RTX 4090 report
Radeon AI PRO R97003-bit32K49 tok/s2.7 sPDF: Qwen2 57B A14B Instruct on Radeon AI PRO R9700 report
RTX 50903-bit32K140 tok/s0.9 sPDF: Qwen2 57B A14B Instruct on RTX 5090 report
2× RTX 30905-bit32K59 tok/s1.5 sPDF: Qwen2 57B A14B Instruct on 2× RTX 3090 report
2× RTX 40905-bit32K63 tok/s0.7 sPDF: Qwen2 57B A14B Instruct on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB6-bit16K72 tok/s1.4 sPDF: Qwen2 57B A14B Instruct on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada6-bit16K53 tok/s1.0 sPDF: Qwen2 57B A14B Instruct on RTX 6000 Ada report
L40S6-bit16K48 tok/s1.6 sPDF: Qwen2 57B A14B Instruct on L40S report
2× RTX 50908-bit32K74 tok/s0.5 sPDF: Qwen2 57B A14B Instruct on 2× RTX 5090 report
4× RTX 30908-bit32K41 tok/s0.8 sPDF: Qwen2 57B A14B Instruct on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit32K74 tok/s0.7 sPDF: Qwen2 57B A14B Instruct on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Qwen2 57B A14B Instruct on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: Qwen2 57B A14B Instruct on Mac mini M4 24 GB report
Mac mini M4 32 GBToo bigPDF: Qwen2 57B A14B Instruct on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: Qwen2 57B A14B Instruct on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB4-bit20K24 tok/s30 sPDF: Qwen2 57B A14B Instruct on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB6-bit20K18 tok/s30 sPDF: Qwen2 57B A14B Instruct on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit32K25 tok/s15 sPDF: Qwen2 57B A14B Instruct on M4 Max 128 GB report
M5 Max 128 GB8-bit32K28 tok/s4.3 sPDF: Qwen2 57B A14B Instruct on M5 Max 128 GB report
M3 Ultra 96 GB8-bit32K31 tok/s8.4 sPDF: Qwen2 57B A14B Instruct on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit32K31 tok/s8.4 sPDF: Qwen2 57B A14B Instruct on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit32K31 tok/s8.4 sPDF: Qwen2 57B A14B Instruct on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit32K43 tok/s2.6 sPDF: Qwen2 57B A14B Instruct on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB6-bit20K13 tok/s8.6 sPDF: Qwen2 57B A14B Instruct on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit32K10 tok/s8.6 sPDF: Qwen2 57B A14B Instruct on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit32K14 tok/s3.0 sPDF: Qwen2 57B A14B Instruct on RTX Spark report
DGX Spark8-bit32K13 tok/s3.0 sPDF: Qwen2 57B A14B Instruct on DGX Spark report
Dual DGX Spark8-bit32K13 tok/s1.7 sPDF: Qwen2 57B A14B Instruct on Dual DGX Spark report
DGX Station GB3008-bit32K214 tok/s0.2 sPDF: Qwen2 57B A14B Instruct on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit32K74 tok/s0.4 sPDF: Qwen2 57B A14B Instruct on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit32K74 tok/s0.2 sPDF: Qwen2 57B A14B Instruct on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit32K74 tok/s0.1 sPDF: Qwen2 57B A14B Instruct on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit32K83 tok/s1.9 sPDF: Qwen2 57B A14B Instruct on A100 80 GB SXM report
H100 SXM8-bit32K125 tok/s0.6 sPDF: Qwen2 57B A14B Instruct on H100 SXM report
4× H200 NVL8-bit32K164 tok/s0.2 sPDF: Qwen2 57B A14B Instruct on 4× H200 NVL report
8× H200 NVL8-bit32K164 tok/s0.1 sPDF: Qwen2 57B A14B Instruct 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 legraphista/Qwen2-57B-A14B-Instruct-GGUF
vLLM
vllm serve Qwen/Qwen2-57B-A14B-Instruct

Questions

How much VRAM does Qwen2 57B A14B Instruct need?

Qwen2 57B A14B Instruct needs 32.4 GB for its weights at 4-bit, 33.7 GB with an 8K context and 35.0 GB with 32K. At 8-bit, 58.1 GB.

What is the minimum VRAM for Qwen2 57B A14B Instruct?

About 27.4 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 Radeon AI PRO R9700.

How big is Qwen2 57B A14B Instruct?

57.4B parameters (14B active per token): 107 GB of weights at 16-bit, 32.4 GB at 4-bit.

Can Qwen2 57B A14B Instruct run on an RTX 4090?

Not on one card. The smallest machine here that runs it is the Radeon AI PRO R9700.

More models