Skip to content

SmolVLM2 2.2B Instruct VRAM requirements

How much VRAM SmolVLM2 2.2B Instruct needs
4-bit weights: 1.3 GBWith 8K context: 1.9 GB8-bit, 8K context: 2.9 GB
Parameters
2.2B
Longest context
8K
Released
Feb 2025
License
Apache 2.0
Strongest at
Text and images

How much VRAM SmolVLM2 2.2B Instruct needs

Memory SmolVLM2 2.2B Instruct needs at each precision and context length
PrecisionWeightsWith 8K
16-bit4.2 GB4.8 GB
8-bit2.2 GB2.9 GB
6-bit1.7 GB2.4 GB
5-bit1.5 GB2.1 GB
4-bit1.3 GB1.9 GB
3-bit1.0 GB1.7 GB
2-bit804 MB1.4 GB
1.6-bit442 MB1.1 GB

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

GPUs and Macs that run SmolVLM2 2.2B Instruct

The best way SmolVLM2 2.2B Instruct runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit8K79 tok/s0.3 sPDF: SmolVLM2 2.2B Instruct on RTX 4060 report
RTX 50608-bit8K125 tok/s0.2 sPDF: SmolVLM2 2.2B Instruct on RTX 5060 report
RTX 30708-bit8K125 tok/s0.2 sPDF: SmolVLM2 2.2B Instruct on RTX 3070 report
RTX 3080 10 GB8-bit8K200 tok/s0.2 sPDF: SmolVLM2 2.2B Instruct on RTX 3080 10 GB report
RTX 3060 12 GB8-bit8K102 tok/s0.4 sPDF: SmolVLM2 2.2B Instruct on RTX 3060 12 GB report
Arc B5808-bit8K87 tok/s0.3 sPDF: SmolVLM2 2.2B Instruct on Arc B580 report
RTX 40708-bit8K139 tok/s0.2 sPDF: SmolVLM2 2.2B Instruct on RTX 4070 report
RTX 4070 Super8-bit8K139 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX 4070 Super report
RTX 50708-bit8K180 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX 5070 report
RTX 4060 Ti 16 GB8-bit8K83 tok/s0.2 sPDF: SmolVLM2 2.2B Instruct on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit8K125 tok/s0.2 sPDF: SmolVLM2 2.2B Instruct on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit8K104 tok/s0.2 sPDF: SmolVLM2 2.2B Instruct on Arc A770 16 GB report
Radeon RX 9070 XT8-bit8K140 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit8K180 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX 4070 Ti Super report
RTX 4080 Super8-bit8K194 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX 4080 Super report
RTX 5070 Ti8-bit8K230 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX 5070 Ti report
RTX 50808-bit8K243 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX 5080 report
Radeon RX 7900 XTX8-bit8K194 tok/s0.2 sPDF: SmolVLM2 2.2B Instruct on Radeon RX 7900 XTX report
RTX 30908-bit8K238 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX 3090 report
RTX 40908-bit8K253 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX 4090 report
Radeon AI PRO R97008-bit8K140 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on Radeon AI PRO R9700 report
RTX 50908-bit8K394 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX 5090 report
2× RTX 30908-bit8K238 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on 2× RTX 3090 report
2× RTX 40908-bit8K253 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit8K318 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit8K243 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX 6000 Ada report
L40S8-bit8K223 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on L40S report
2× RTX 50908-bit8K394 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on 2× RTX 5090 report
4× RTX 30908-bit8K238 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit8K394 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit8K40 tok/s2.9 sPDF: SmolVLM2 2.2B Instruct on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit8K40 tok/s2.9 sPDF: SmolVLM2 2.2B Instruct on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit8K40 tok/s2.9 sPDF: SmolVLM2 2.2B Instruct on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit8K83 tok/s1.5 sPDF: SmolVLM2 2.2B Instruct on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit8K83 tok/s1.5 sPDF: SmolVLM2 2.2B Instruct on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit8K83 tok/s1.5 sPDF: SmolVLM2 2.2B Instruct on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit8K134 tok/s0.7 sPDF: SmolVLM2 2.2B Instruct on M4 Max 128 GB report
M5 Max 128 GB8-bit8K146 tok/s0.2 sPDF: SmolVLM2 2.2B Instruct on M5 Max 128 GB report
M3 Ultra 96 GB8-bit8K161 tok/s0.4 sPDF: SmolVLM2 2.2B Instruct on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit8K161 tok/s0.4 sPDF: SmolVLM2 2.2B Instruct on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit8K161 tok/s0.4 sPDF: SmolVLM2 2.2B Instruct on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit8K206 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit8K62 tok/s0.4 sPDF: SmolVLM2 2.2B Instruct on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit8K62 tok/s0.4 sPDF: SmolVLM2 2.2B Instruct on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit8K86 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on RTX Spark report
DGX Spark8-bit8K79 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on DGX Spark report
Dual DGX Spark8-bit8K79 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on Dual DGX Spark report
DGX Station GB3008-bit8K848 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit8K394 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit8K394 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit8K394 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit8K432 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on A100 80 GB SXM report
H100 SXM8-bit8K591 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on H100 SXM report
4× H200 NVL8-bit8K715 tok/s0.1 sPDF: SmolVLM2 2.2B Instruct on 4× H200 NVL report
8× H200 NVL8-bit8K715 tok/s0.1 sPDF: SmolVLM2 2.2B 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 ggml-org/SmolVLM2-2.2B-Instruct-GGUF
vLLM
vllm serve HuggingFaceTB/SmolVLM2-2.2B-Instruct

Questions

How much VRAM does SmolVLM2 2.2B Instruct need?

SmolVLM2 2.2B Instruct needs 1.3 GB for its weights at 4-bit, 1.9 GB with an 8K context and 1.9 GB with 8K. At 8-bit, 2.9 GB.

What is the minimum VRAM for SmolVLM2 2.2B Instruct?

About 1.7 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 SmolVLM2 2.2B Instruct?

2.2B parameters: 4.2 GB of weights at 16-bit, 1.3 GB at 4-bit.

Can SmolVLM2 2.2B Instruct run on an RTX 4090?

Yes. At 8-bit it holds 8K of context and writes about 253 tokens a second.

More models