Skip to content

Qwen2.5 14B Instruct 1M VRAM requirements

How much VRAM Qwen2.5 14B Instruct 1M needs
4-bit weights: 8.3 GBWith 8K context: 10.7 GBWith 32K context: 15.2 GB8-bit, 8K context: 17.0 GB
Parameters
14.8B
Longest context
986K
Released
Jan 2025
License
Apache 2.0
Strongest at
Not rated

How much VRAM Qwen2.5 14B Instruct 1M needs

Memory Qwen2.5 14B Instruct 1M needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit27.5 GB29.9 GB34.4 GB52.4 GB
8-bit14.6 GB17.0 GB21.5 GB39.5 GB
6-bit11.3 GB13.7 GB18.2 GB36.2 GB
5-bit9.8 GB12.2 GB16.7 GB34.7 GB
4-bit8.3 GB10.7 GB15.2 GB33.2 GB
3-bit6.7 GB9.1 GB13.6 GB31.6 GB
2-bit5.2 GB7.6 GB12.1 GB30.1 GB
1.6-bit2.8 GB5.2 GB9.7 GB27.7 GB

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

GPUs and Macs that run Qwen2.5 14B Instruct 1M

The best way Qwen2.5 14B Instruct 1M runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Qwen2.5 14B Instruct 1M on RTX 4060 report
RTX 5060Too bigPDF: Qwen2.5 14B Instruct 1M on RTX 5060 report
RTX 3070Too bigPDF: Qwen2.5 14B Instruct 1M on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Qwen2.5 14B Instruct 1M on RTX 3080 10 GB report
RTX 3060 12 GB3-bit20K35 tok/s2.3 sPDF: Qwen2.5 14B Instruct 1M on RTX 3060 12 GB report
Arc B5803-bit20K30 tok/s1.6 sPDF: Qwen2.5 14B Instruct 1M on Arc B580 report
RTX 40703-bit20K48 tok/s1.0 sPDF: Qwen2.5 14B Instruct 1M on RTX 4070 report
RTX 4070 Super3-bit20K48 tok/s0.8 sPDF: Qwen2.5 14B Instruct 1M on RTX 4070 Super report
RTX 50703-bit20K63 tok/s1.0 sPDF: Qwen2.5 14B Instruct 1M on RTX 5070 report
RTX 4060 Ti 16 GB5-bit20K20 tok/s1.3 sPDF: Qwen2.5 14B Instruct 1M on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB5-bit20K30 tok/s1.2 sPDF: Qwen2.5 14B Instruct 1M on RTX 5060 Ti 16 GB report
Arc A770 16 GB5-bit20K26 tok/s1.4 sPDF: Qwen2.5 14B Instruct 1M on Arc A770 16 GB report
Radeon RX 9070 XT5-bit20K35 tok/s0.8 sPDF: Qwen2.5 14B Instruct 1M on Radeon RX 9070 XT report
RTX 4070 Ti Super5-bit20K44 tok/s0.7 sPDF: Qwen2.5 14B Instruct 1M on RTX 4070 Ti Super report
RTX 4080 Super5-bit20K48 tok/s0.6 sPDF: Qwen2.5 14B Instruct 1M on RTX 4080 Super report
RTX 5070 Ti5-bit20K58 tok/s0.7 sPDF: Qwen2.5 14B Instruct 1M on RTX 5070 Ti report
RTX 50805-bit20K62 tok/s0.5 sPDF: Qwen2.5 14B Instruct 1M on RTX 5080 report
Radeon RX 7900 XTX8-bit36K35 tok/s1.3 sPDF: Qwen2.5 14B Instruct 1M on Radeon RX 7900 XTX report
RTX 30908-bit36K41 tok/s0.8 sPDF: Qwen2.5 14B Instruct 1M on RTX 3090 report
RTX 40908-bit36K44 tok/s0.4 sPDF: Qwen2.5 14B Instruct 1M on RTX 4090 report
Radeon AI PRO R97008-bit72K24 tok/s0.8 sPDF: Qwen2.5 14B Instruct 1M on Radeon AI PRO R9700 report
RTX 50908-bit72K75 tok/s0.3 sPDF: Qwen2.5 14B Instruct 1M on RTX 5090 report
2× RTX 30908-bit152K41 tok/s0.5 sPDF: Qwen2.5 14B Instruct 1M on 2× RTX 3090 report
2× RTX 40908-bit152K44 tok/s0.2 sPDF: Qwen2.5 14B Instruct 1M on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit160K58 tok/s0.4 sPDF: Qwen2.5 14B Instruct 1M on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit160K42 tok/s0.3 sPDF: Qwen2.5 14B Instruct 1M on RTX 6000 Ada report
L40S8-bit160K38 tok/s0.5 sPDF: Qwen2.5 14B Instruct 1M on L40S report
2× RTX 50908-bit232K75 tok/s0.2 sPDF: Qwen2.5 14B Instruct 1M on 2× RTX 5090 report
4× RTX 30908-bit384K41 tok/s0.3 sPDF: Qwen2.5 14B Instruct 1M on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit384K75 tok/s0.2 sPDF: Qwen2.5 14B Instruct 1M on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Qwen2.5 14B Instruct 1M on Mac mini M4 16 GB report
Mac mini M4 24 GB6-bit16K8.3 tok/s19 sPDF: Qwen2.5 14B Instruct 1M on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit24K6.5 tok/s19 sPDF: Qwen2.5 14B Instruct 1M on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB6-bit16K18 tok/s9.4 sPDF: Qwen2.5 14B Instruct 1M on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit96K14 tok/s9.4 sPDF: Qwen2.5 14B Instruct 1M on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit160K14 tok/s9.4 sPDF: Qwen2.5 14B Instruct 1M on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit384K25 tok/s4.7 sPDF: Qwen2.5 14B Instruct 1M on M4 Max 128 GB report
M5 Max 128 GB8-bit384K28 tok/s1.4 sPDF: Qwen2.5 14B Instruct 1M on M5 Max 128 GB report
M3 Ultra 96 GB8-bit256K31 tok/s2.6 sPDF: Qwen2.5 14B Instruct 1M on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit864K31 tok/s2.6 sPDF: Qwen2.5 14B Instruct 1M on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit986K31 tok/s2.6 sPDF: Qwen2.5 14B Instruct 1M on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit986K44 tok/s0.8 sPDF: Qwen2.5 14B Instruct 1M on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit160K10 tok/s2.7 sPDF: Qwen2.5 14B Instruct 1M on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit384K10 tok/s2.7 sPDF: Qwen2.5 14B Instruct 1M on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit512K14 tok/s0.9 sPDF: Qwen2.5 14B Instruct 1M on RTX Spark report
DGX Spark8-bit512K13 tok/s0.9 sPDF: Qwen2.5 14B Instruct 1M on DGX Spark report
Dual DGX Spark8-bit986K13 tok/s0.5 sPDF: Qwen2.5 14B Instruct 1M on Dual DGX Spark report
DGX Station GB3008-bit986K226 tok/s0.1 sPDF: Qwen2.5 14B Instruct 1M on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit864K75 tok/s0.1 sPDF: Qwen2.5 14B Instruct 1M on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit986K75 tok/s0.1 sPDF: Qwen2.5 14B Instruct 1M on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit986K75 tok/s0.1 sPDF: Qwen2.5 14B Instruct 1M on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit320K84 tok/s0.6 sPDF: Qwen2.5 14B Instruct 1M on A100 80 GB SXM report
H100 SXM8-bit320K129 tok/s0.2 sPDF: Qwen2.5 14B Instruct 1M on H100 SXM report
4× H200 NVL8-bit986K171 tok/s0.1 sPDF: Qwen2.5 14B Instruct 1M on 4× H200 NVL report
8× H200 NVL8-bit986K171 tok/s0.1 sPDF: Qwen2.5 14B Instruct 1M 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 lmstudio-community/Qwen2.5-14B-Instruct-1M-GGUF
vLLM
vllm serve Qwen/Qwen2.5-14B-Instruct-1M
MLX (Mac)
mlx_lm.chat --model mlx-community/Qwen2.5-14B-Instruct-1M-8bit

Questions

How much VRAM does Qwen2.5 14B Instruct 1M need?

Qwen2.5 14B Instruct 1M needs 8.3 GB for its weights at 4-bit, 10.7 GB with an 8K context and 15.2 GB with 32K. At 8-bit, 17.0 GB.

What is the minimum VRAM for Qwen2.5 14B Instruct 1M?

About 9.1 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 3060 12 GB.

How big is Qwen2.5 14B Instruct 1M?

14.8B parameters: 27.5 GB of weights at 16-bit, 8.3 GB at 4-bit.

Can Qwen2.5 14B Instruct 1M run on an RTX 4090?

Yes. At 8-bit it holds 36K of context and writes about 44 tokens a second.

More models