Skip to content

Stablelm 3B 4e1t VRAM requirements

How much VRAM Stablelm 3B 4e1t needs
4-bit weights: 1.6 GBWith 4K context: 3.5 GB8-bit, 4K context: 4.6 GB
Parameters
2.8B
Longest context
4K
Released
Sep 2023
License
CC BY-SA 4.0
Strongest at
Not rated

How much VRAM Stablelm 3B 4e1t needs

Memory Stablelm 3B 4e1t needs at each precision and context length
PrecisionWeightsWith 4K
16-bit5.2 GB7.1 GB
8-bit2.8 GB4.6 GB
6-bit2.1 GB4.0 GB
5-bit1.9 GB3.7 GB
4-bit1.6 GB3.5 GB
3-bit1.3 GB3.1 GB
2-bit1.0 GB2.9 GB
1.6-bit550 MB2.4 GB

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

GPUs and Macs that run Stablelm 3B 4e1t

The best way Stablelm 3B 4e1t runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit4K63 tok/s0.4 sPDF: Stablelm 3B 4e1t on RTX 4060 report
RTX 50608-bit4K100 tok/s0.3 sPDF: Stablelm 3B 4e1t on RTX 5060 report
RTX 30708-bit4K100 tok/s0.3 sPDF: Stablelm 3B 4e1t on RTX 3070 report
RTX 3080 10 GB8-bit4K159 tok/s0.2 sPDF: Stablelm 3B 4e1t on RTX 3080 10 GB report
RTX 3060 12 GB8-bit4K82 tok/s0.5 sPDF: Stablelm 3B 4e1t on RTX 3060 12 GB report
Arc B5808-bit4K70 tok/s0.3 sPDF: Stablelm 3B 4e1t on Arc B580 report
RTX 40708-bit4K111 tok/s0.2 sPDF: Stablelm 3B 4e1t on RTX 4070 report
RTX 4070 Super8-bit4K111 tok/s0.2 sPDF: Stablelm 3B 4e1t on RTX 4070 Super report
RTX 50708-bit4K143 tok/s0.2 sPDF: Stablelm 3B 4e1t on RTX 5070 report
RTX 4060 Ti 16 GB8-bit4K67 tok/s0.3 sPDF: Stablelm 3B 4e1t on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit4K100 tok/s0.2 sPDF: Stablelm 3B 4e1t on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit4K83 tok/s0.3 sPDF: Stablelm 3B 4e1t on Arc A770 16 GB report
Radeon RX 9070 XT8-bit4K111 tok/s0.2 sPDF: Stablelm 3B 4e1t on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit4K143 tok/s0.1 sPDF: Stablelm 3B 4e1t on RTX 4070 Ti Super report
RTX 4080 Super8-bit4K155 tok/s0.1 sPDF: Stablelm 3B 4e1t on RTX 4080 Super report
RTX 5070 Ti8-bit4K183 tok/s0.1 sPDF: Stablelm 3B 4e1t on RTX 5070 Ti report
RTX 50808-bit4K193 tok/s0.1 sPDF: Stablelm 3B 4e1t on RTX 5080 report
Radeon RX 7900 XTX8-bit4K153 tok/s0.3 sPDF: Stablelm 3B 4e1t on Radeon RX 7900 XTX report
RTX 30908-bit4K189 tok/s0.2 sPDF: Stablelm 3B 4e1t on RTX 3090 report
RTX 40908-bit4K201 tok/s0.1 sPDF: Stablelm 3B 4e1t on RTX 4090 report
Radeon AI PRO R97008-bit4K111 tok/s0.2 sPDF: Stablelm 3B 4e1t on Radeon AI PRO R9700 report
RTX 50908-bit4K311 tok/s0.1 sPDF: Stablelm 3B 4e1t on RTX 5090 report
2× RTX 30908-bit4K189 tok/s0.1 sPDF: Stablelm 3B 4e1t on 2× RTX 3090 report
2× RTX 40908-bit4K201 tok/s0.1 sPDF: Stablelm 3B 4e1t on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit4K252 tok/s0.1 sPDF: Stablelm 3B 4e1t on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit4K193 tok/s0.1 sPDF: Stablelm 3B 4e1t on RTX 6000 Ada report
L40S8-bit4K177 tok/s0.1 sPDF: Stablelm 3B 4e1t on L40S report
2× RTX 50908-bit4K311 tok/s0.1 sPDF: Stablelm 3B 4e1t on 2× RTX 5090 report
4× RTX 30908-bit4K189 tok/s0.1 sPDF: Stablelm 3B 4e1t on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit4K311 tok/s0.1 sPDF: Stablelm 3B 4e1t on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit4K32 tok/s3.7 sPDF: Stablelm 3B 4e1t on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit4K32 tok/s3.7 sPDF: Stablelm 3B 4e1t on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit4K32 tok/s3.7 sPDF: Stablelm 3B 4e1t on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit4K66 tok/s1.9 sPDF: Stablelm 3B 4e1t on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit4K66 tok/s1.9 sPDF: Stablelm 3B 4e1t on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit4K66 tok/s1.9 sPDF: Stablelm 3B 4e1t on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit4K105 tok/s0.9 sPDF: Stablelm 3B 4e1t on M4 Max 128 GB report
M5 Max 128 GB8-bit4K115 tok/s0.3 sPDF: Stablelm 3B 4e1t on M5 Max 128 GB report
M3 Ultra 96 GB8-bit4K127 tok/s0.5 sPDF: Stablelm 3B 4e1t on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit4K127 tok/s0.5 sPDF: Stablelm 3B 4e1t on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit4K127 tok/s0.5 sPDF: Stablelm 3B 4e1t on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit4K161 tok/s0.2 sPDF: Stablelm 3B 4e1t on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit4K50 tok/s0.5 sPDF: Stablelm 3B 4e1t on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit4K50 tok/s0.5 sPDF: Stablelm 3B 4e1t on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit4K69 tok/s0.2 sPDF: Stablelm 3B 4e1t on RTX Spark report
DGX Spark8-bit4K63 tok/s0.2 sPDF: Stablelm 3B 4e1t on DGX Spark report
Dual DGX Spark8-bit4K63 tok/s0.1 sPDF: Stablelm 3B 4e1t on Dual DGX Spark report
DGX Station GB3008-bit4K654 tok/s0.1 sPDF: Stablelm 3B 4e1t on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit4K311 tok/s0.1 sPDF: Stablelm 3B 4e1t on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit4K311 tok/s0.1 sPDF: Stablelm 3B 4e1t on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit4K311 tok/s0.1 sPDF: Stablelm 3B 4e1t on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit4K340 tok/s0.1 sPDF: Stablelm 3B 4e1t on A100 80 GB SXM report
H100 SXM8-bit4K461 tok/s0.1 sPDF: Stablelm 3B 4e1t on H100 SXM report
4× H200 NVL8-bit4K555 tok/s0.1 sPDF: Stablelm 3B 4e1t on 4× H200 NVL report
8× H200 NVL8-bit4K555 tok/s0.1 sPDF: Stablelm 3B 4e1t 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 afrideva/stablelm-3b-4e1t-GGUF
vLLM
vllm serve stabilityai/stablelm-3b-4e1t

Questions

How much VRAM does Stablelm 3B 4e1t need?

Stablelm 3B 4e1t needs 1.6 GB for its weights at 4-bit, 3.5 GB with an 4K context and 3.5 GB with 4K. At 8-bit, 4.6 GB.

What is the minimum VRAM for Stablelm 3B 4e1t?

About 3.1 GB at 3-bit with an 4K 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 Stablelm 3B 4e1t?

2.8B parameters: 5.2 GB of weights at 16-bit, 1.6 GB at 4-bit.

Can Stablelm 3B 4e1t run on an RTX 4090?

Yes. At 8-bit it holds 4K of context and writes about 201 tokens a second.

More models