Skip to content

Qwen2.5 7B Instruct 1M VRAM requirements

How much VRAM Qwen2.5 7B Instruct 1M needs
4-bit weights: 4.3 GBWith 8K context: 5.6 GBWith 32K context: 6.9 GB8-bit, 8K context: 8.8 GB
Parameters
7.6B
Longest context
986K
Released
Jan 2025
License
Apache 2.0
Strongest at
Not rated

How much VRAM Qwen2.5 7B Instruct 1M needs

Memory Qwen2.5 7B Instruct 1M needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit14.2 GB15.5 GB16.8 GB22.0 GB
8-bit7.5 GB8.8 GB10.1 GB15.4 GB
6-bit5.8 GB7.1 GB8.4 GB13.7 GB
5-bit5.0 GB6.3 GB7.6 GB12.9 GB
4-bit4.3 GB5.6 GB6.9 GB12.1 GB
3-bit3.5 GB4.8 GB6.1 GB11.3 GB
2-bit2.7 GB3.9 GB5.3 GB10.5 GB
1.6-bit1.5 GB2.7 GB4.1 GB9.3 GB

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

GPUs and Macs that run Qwen2.5 7B Instruct 1M

The best way Qwen2.5 7B Instruct 1M runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40606-bit16K32 tok/s1.0 sPDF: Qwen2.5 7B Instruct 1M on RTX 4060 report
RTX 50606-bit16K52 tok/s0.8 sPDF: Qwen2.5 7B Instruct 1M on RTX 5060 report
RTX 30706-bit16K52 tok/s0.7 sPDF: Qwen2.5 7B Instruct 1M on RTX 3070 report
RTX 3080 10 GB8-bit20K67 tok/s0.5 sPDF: Qwen2.5 7B Instruct 1M on RTX 3080 10 GB report
RTX 3060 12 GB8-bit52K33 tok/s1.1 sPDF: Qwen2.5 7B Instruct 1M on RTX 3060 12 GB report
Arc B5808-bit52K29 tok/s0.8 sPDF: Qwen2.5 7B Instruct 1M on Arc B580 report
RTX 40708-bit52K45 tok/s0.5 sPDF: Qwen2.5 7B Instruct 1M on RTX 4070 report
RTX 4070 Super8-bit52K45 tok/s0.4 sPDF: Qwen2.5 7B Instruct 1M on RTX 4070 Super report
RTX 50708-bit52K59 tok/s0.5 sPDF: Qwen2.5 7B Instruct 1M on RTX 5070 report
RTX 4060 Ti 16 GB8-bit120K26 tok/s0.7 sPDF: Qwen2.5 7B Instruct 1M on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit120K40 tok/s0.6 sPDF: Qwen2.5 7B Instruct 1M on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit120K35 tok/s0.7 sPDF: Qwen2.5 7B Instruct 1M on Arc A770 16 GB report
Radeon RX 9070 XT8-bit120K48 tok/s0.4 sPDF: Qwen2.5 7B Instruct 1M on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit120K59 tok/s0.3 sPDF: Qwen2.5 7B Instruct 1M on RTX 4070 Ti Super report
RTX 4080 Super8-bit120K65 tok/s0.3 sPDF: Qwen2.5 7B Instruct 1M on RTX 4080 Super report
RTX 5070 Ti8-bit120K78 tok/s0.3 sPDF: Qwen2.5 7B Instruct 1M on RTX 5070 Ti report
RTX 50808-bit120K83 tok/s0.3 sPDF: Qwen2.5 7B Instruct 1M on RTX 5080 report
Radeon RX 7900 XTX8-bit256K69 tok/s0.6 sPDF: Qwen2.5 7B Instruct 1M on Radeon RX 7900 XTX report
RTX 30908-bit256K81 tok/s0.4 sPDF: Qwen2.5 7B Instruct 1M on RTX 3090 report
RTX 40908-bit256K87 tok/s0.2 sPDF: Qwen2.5 7B Instruct 1M on RTX 4090 report
Radeon AI PRO R97008-bit384K48 tok/s0.4 sPDF: Qwen2.5 7B Instruct 1M on Radeon AI PRO R9700 report
RTX 50908-bit384K146 tok/s0.1 sPDF: Qwen2.5 7B Instruct 1M on RTX 5090 report
2× RTX 30908-bit640K81 tok/s0.2 sPDF: Qwen2.5 7B Instruct 1M on 2× RTX 3090 report
2× RTX 40908-bit640K87 tok/s0.1 sPDF: Qwen2.5 7B Instruct 1M on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit672K113 tok/s0.2 sPDF: Qwen2.5 7B Instruct 1M on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit672K83 tok/s0.2 sPDF: Qwen2.5 7B Instruct 1M on RTX 6000 Ada report
L40S8-bit672K75 tok/s0.3 sPDF: Qwen2.5 7B Instruct 1M on L40S report
2× RTX 50908-bit928K146 tok/s0.1 sPDF: Qwen2.5 7B Instruct 1M on 2× RTX 5090 report
4× RTX 30908-bit986K81 tok/s0.1 sPDF: Qwen2.5 7B Instruct 1M on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit986K146 tok/s0.1 sPDF: Qwen2.5 7B Instruct 1M on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit32K13 tok/s9.2 sPDF: Qwen2.5 7B Instruct 1M on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit128K13 tok/s9.2 sPDF: Qwen2.5 7B Instruct 1M on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit216K13 tok/s9.2 sPDF: Qwen2.5 7B Instruct 1M on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit128K28 tok/s4.6 sPDF: Qwen2.5 7B Instruct 1M on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit448K28 tok/s4.6 sPDF: Qwen2.5 7B Instruct 1M on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit672K28 tok/s4.6 sPDF: Qwen2.5 7B Instruct 1M on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit986K49 tok/s2.3 sPDF: Qwen2.5 7B Instruct 1M on M4 Max 128 GB report
M5 Max 128 GB8-bit986K54 tok/s0.7 sPDF: Qwen2.5 7B Instruct 1M on M5 Max 128 GB report
M3 Ultra 96 GB8-bit986K61 tok/s1.3 sPDF: Qwen2.5 7B Instruct 1M on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit986K61 tok/s1.3 sPDF: Qwen2.5 7B Instruct 1M on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit986K61 tok/s1.3 sPDF: Qwen2.5 7B Instruct 1M on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit986K84 tok/s0.4 sPDF: Qwen2.5 7B Instruct 1M on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit672K20 tok/s1.3 sPDF: Qwen2.5 7B Instruct 1M on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit986K20 tok/s1.3 sPDF: Qwen2.5 7B Instruct 1M on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit986K27 tok/s0.5 sPDF: Qwen2.5 7B Instruct 1M on RTX Spark report
DGX Spark8-bit986K25 tok/s0.5 sPDF: Qwen2.5 7B Instruct 1M on DGX Spark report
Dual DGX Spark8-bit986K25 tok/s0.3 sPDF: Qwen2.5 7B Instruct 1M on Dual DGX Spark report
DGX Station GB3008-bit986K425 tok/s0.1 sPDF: Qwen2.5 7B Instruct 1M on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit986K146 tok/s0.1 sPDF: Qwen2.5 7B Instruct 1M on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit986K146 tok/s0.1 sPDF: Qwen2.5 7B Instruct 1M on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit986K146 tok/s0.1 sPDF: Qwen2.5 7B Instruct 1M on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit986K164 tok/s0.3 sPDF: Qwen2.5 7B Instruct 1M on A100 80 GB SXM report
H100 SXM8-bit986K247 tok/s0.1 sPDF: Qwen2.5 7B Instruct 1M on H100 SXM report
4× H200 NVL8-bit986K325 tok/s0.1 sPDF: Qwen2.5 7B Instruct 1M on 4× H200 NVL report
8× H200 NVL8-bit986K325 tok/s0.1 sPDF: Qwen2.5 7B 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/Qwen2.5-7B-Instruct-1M-GGUF
vLLM
vllm serve Qwen/Qwen2.5-7B-Instruct-1M
MLX (Mac)
mlx_lm.chat --model mlx-community/Qwen2.5-7B-Instruct-1M-4bit

Questions

How much VRAM does Qwen2.5 7B Instruct 1M need?

Qwen2.5 7B Instruct 1M needs 4.3 GB for its weights at 4-bit, 5.6 GB with an 8K context and 6.9 GB with 32K. At 8-bit, 8.8 GB.

What is the minimum VRAM for Qwen2.5 7B Instruct 1M?

About 4.8 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 Qwen2.5 7B Instruct 1M?

7.6B parameters: 14.2 GB of weights at 16-bit, 4.3 GB at 4-bit.

Can Qwen2.5 7B Instruct 1M run on an RTX 4090?

Yes. At 8-bit it holds 256K of context and writes about 87 tokens a second.

More models