Skip to content

Qwen2.5 Coder 3B Instruct VRAM requirements

How much VRAM Qwen2.5 Coder 3B Instruct needs
4-bit weights: 1.7 GBWith 8K context: 2.8 GBWith 32K context: 3.7 GB8-bit, 8K context: 4.1 GB
Parameters
3.1B
Longest context
32K
Released
Nov 2024
License
Qwen Research
Strongest at
Coding

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

How much VRAM Qwen2.5 Coder 3B Instruct needs

Memory Qwen2.5 Coder 3B Instruct needs at each precision and context length
PrecisionWeightsWith 8KWith 32K
16-bit5.7 GB6.8 GB7.7 GB
8-bit3.1 GB4.1 GB5.0 GB
6-bit2.4 GB3.4 GB4.3 GB
5-bit2.0 GB3.1 GB4.0 GB
4-bit1.7 GB2.8 GB3.7 GB
3-bit1.4 GB2.5 GB3.3 GB
2-bit1.1 GB2.2 GB3.0 GB
1.6-bit607 MB1.7 GB2.5 GB

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

GPUs and Macs that run Qwen2.5 Coder 3B Instruct

The best way Qwen2.5 Coder 3B Instruct runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit32K55 tok/s0.4 sPDF: Qwen2.5 Coder 3B Instruct on RTX 4060 report
RTX 50608-bit32K87 tok/s0.3 sPDF: Qwen2.5 Coder 3B Instruct on RTX 5060 report
RTX 30708-bit32K87 tok/s0.3 sPDF: Qwen2.5 Coder 3B Instruct on RTX 3070 report
RTX 3080 10 GB8-bit32K138 tok/s0.2 sPDF: Qwen2.5 Coder 3B Instruct on RTX 3080 10 GB report
RTX 3060 12 GB8-bit32K71 tok/s0.5 sPDF: Qwen2.5 Coder 3B Instruct on RTX 3060 12 GB report
Arc B5808-bit32K60 tok/s0.3 sPDF: Qwen2.5 Coder 3B Instruct on Arc B580 report
RTX 40708-bit32K96 tok/s0.2 sPDF: Qwen2.5 Coder 3B Instruct on RTX 4070 report
RTX 4070 Super8-bit32K96 tok/s0.2 sPDF: Qwen2.5 Coder 3B Instruct on RTX 4070 Super report
RTX 50708-bit32K124 tok/s0.2 sPDF: Qwen2.5 Coder 3B Instruct on RTX 5070 report
RTX 4060 Ti 16 GB8-bit32K58 tok/s0.3 sPDF: Qwen2.5 Coder 3B Instruct on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit32K87 tok/s0.3 sPDF: Qwen2.5 Coder 3B Instruct on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit32K72 tok/s0.3 sPDF: Qwen2.5 Coder 3B Instruct on Arc A770 16 GB report
Radeon RX 9070 XT8-bit32K97 tok/s0.2 sPDF: Qwen2.5 Coder 3B Instruct on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit32K124 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on RTX 4070 Ti Super report
RTX 4080 Super8-bit32K134 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on RTX 4080 Super report
RTX 5070 Ti8-bit32K159 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on RTX 5070 Ti report
RTX 50808-bit32K168 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on RTX 5080 report
Radeon RX 7900 XTX8-bit32K133 tok/s0.3 sPDF: Qwen2.5 Coder 3B Instruct on Radeon RX 7900 XTX report
RTX 30908-bit32K164 tok/s0.2 sPDF: Qwen2.5 Coder 3B Instruct on RTX 3090 report
RTX 40908-bit32K175 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on RTX 4090 report
Radeon AI PRO R97008-bit32K97 tok/s0.2 sPDF: Qwen2.5 Coder 3B Instruct on Radeon AI PRO R9700 report
RTX 50908-bit32K271 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on RTX 5090 report
2× RTX 30908-bit32K164 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on 2× RTX 3090 report
2× RTX 40908-bit32K175 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit32K219 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit32K168 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on RTX 6000 Ada report
L40S8-bit32K154 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on L40S report
2× RTX 50908-bit32K271 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on 2× RTX 5090 report
4× RTX 30908-bit32K164 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit32K271 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit32K28 tok/s4.0 sPDF: Qwen2.5 Coder 3B Instruct on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit32K28 tok/s4.0 sPDF: Qwen2.5 Coder 3B Instruct on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit32K28 tok/s4.0 sPDF: Qwen2.5 Coder 3B Instruct on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit32K57 tok/s2.0 sPDF: Qwen2.5 Coder 3B Instruct on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit32K57 tok/s2.0 sPDF: Qwen2.5 Coder 3B Instruct on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit32K57 tok/s2.0 sPDF: Qwen2.5 Coder 3B Instruct on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit32K92 tok/s1.0 sPDF: Qwen2.5 Coder 3B Instruct on M4 Max 128 GB report
M5 Max 128 GB8-bit32K100 tok/s0.3 sPDF: Qwen2.5 Coder 3B Instruct on M5 Max 128 GB report
M3 Ultra 96 GB8-bit32K111 tok/s0.6 sPDF: Qwen2.5 Coder 3B Instruct on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit32K111 tok/s0.6 sPDF: Qwen2.5 Coder 3B Instruct on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit32K111 tok/s0.6 sPDF: Qwen2.5 Coder 3B Instruct on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit32K141 tok/s0.2 sPDF: Qwen2.5 Coder 3B Instruct on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit32K43 tok/s0.6 sPDF: Qwen2.5 Coder 3B Instruct on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit32K43 tok/s0.6 sPDF: Qwen2.5 Coder 3B Instruct on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit32K60 tok/s0.2 sPDF: Qwen2.5 Coder 3B Instruct on RTX Spark report
DGX Spark8-bit32K55 tok/s0.2 sPDF: Qwen2.5 Coder 3B Instruct on DGX Spark report
Dual DGX Spark8-bit32K55 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on Dual DGX Spark report
DGX Station GB3008-bit32K575 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit32K271 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit32K271 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit32K271 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit32K296 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on A100 80 GB SXM report
H100 SXM8-bit32K404 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on H100 SXM report
4× H200 NVL8-bit32K486 tok/s0.1 sPDF: Qwen2.5 Coder 3B Instruct on 4× H200 NVL report
8× H200 NVL8-bit32K486 tok/s0.1 sPDF: Qwen2.5 Coder 3B 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 Qwen/Qwen2.5-Coder-3B-Instruct-GGUF
vLLM
vllm serve Qwen/Qwen2.5-Coder-3B-Instruct
MLX (Mac)
mlx_lm.chat --model mlx-community/Qwen2.5-Coder-3B-Instruct-4bit

Questions

How much VRAM does Qwen2.5 Coder 3B Instruct need?

Qwen2.5 Coder 3B Instruct needs 1.7 GB for its weights at 4-bit, 2.8 GB with an 8K context and 3.7 GB with 32K. At 8-bit, 4.1 GB.

What is the minimum VRAM for Qwen2.5 Coder 3B Instruct?

About 2.5 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 Coder 3B Instruct?

3.1B parameters: 5.7 GB of weights at 16-bit, 1.7 GB at 4-bit.

Can Qwen2.5 Coder 3B Instruct run on an RTX 4090?

Yes. At 8-bit it holds 32K of context and writes about 175 tokens a second.

More models