Skip to content

OLMoE 1B 7B 0924 VRAM requirements

How much VRAM OLMoE 1B 7B 0924 needs
4-bit weights: 3.9 GBWith 4K context: 5.0 GB8-bit, 4K context: 8.0 GB
Parameters
6.9B
Active per token
1.3B
Longest context
4K
Released
Jul 2024
License
Apache 2.0
Strongest at
Not rated

How much VRAM OLMoE 1B 7B 0924 needs

Memory OLMoE 1B 7B 0924 needs at each precision and context length
PrecisionWeightsWith 4K
16-bit12.9 GB14.0 GB
8-bit6.8 GB8.0 GB
6-bit5.3 GB6.4 GB
5-bit4.6 GB5.7 GB
4-bit3.9 GB5.0 GB
3-bit3.1 GB4.3 GB
2-bit2.4 GB3.5 GB
1.6-bit1.3 GB2.4 GB

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

GPUs and Macs that run OLMoE 1B 7B 0924

The best way OLMoE 1B 7B 0924 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit4K123 tok/s0.5 sPDF: OLMoE 1B 7B 0924 on RTX 4060 report
RTX 50608-bit4K189 tok/s0.4 sPDF: OLMoE 1B 7B 0924 on RTX 5060 report
RTX 30708-bit4K189 tok/s0.4 sPDF: OLMoE 1B 7B 0924 on RTX 3070 report
RTX 3080 10 GB8-bit4K284 tok/s0.3 sPDF: OLMoE 1B 7B 0924 on RTX 3080 10 GB report
RTX 3060 12 GB8-bit4K157 tok/s0.6 sPDF: OLMoE 1B 7B 0924 on RTX 3060 12 GB report
Arc B5808-bit4K132 tok/s0.4 sPDF: OLMoE 1B 7B 0924 on Arc B580 report
RTX 40708-bit4K207 tok/s0.3 sPDF: OLMoE 1B 7B 0924 on RTX 4070 report
RTX 4070 Super8-bit4K207 tok/s0.2 sPDF: OLMoE 1B 7B 0924 on RTX 4070 Super report
RTX 50708-bit4K259 tok/s0.3 sPDF: OLMoE 1B 7B 0924 on RTX 5070 report
RTX 4060 Ti 16 GB8-bit4K130 tok/s0.4 sPDF: OLMoE 1B 7B 0924 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit4K189 tok/s0.3 sPDF: OLMoE 1B 7B 0924 on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit4K155 tok/s0.4 sPDF: OLMoE 1B 7B 0924 on Arc A770 16 GB report
Radeon RX 9070 XT8-bit4K201 tok/s0.2 sPDF: OLMoE 1B 7B 0924 on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit4K259 tok/s0.2 sPDF: OLMoE 1B 7B 0924 on RTX 4070 Ti Super report
RTX 4080 Super8-bit4K277 tok/s0.2 sPDF: OLMoE 1B 7B 0924 on RTX 4080 Super report
RTX 5070 Ti8-bit4K319 tok/s0.2 sPDF: OLMoE 1B 7B 0924 on RTX 5070 Ti report
RTX 50808-bit4K335 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on RTX 5080 report
Radeon RX 7900 XTX8-bit4K263 tok/s0.4 sPDF: OLMoE 1B 7B 0924 on Radeon RX 7900 XTX report
RTX 30908-bit4K329 tok/s0.2 sPDF: OLMoE 1B 7B 0924 on RTX 3090 report
RTX 40908-bit4K346 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on RTX 4090 report
Radeon AI PRO R97008-bit4K201 tok/s0.2 sPDF: OLMoE 1B 7B 0924 on Radeon AI PRO R9700 report
RTX 50908-bit4K489 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on RTX 5090 report
2× RTX 30908-bit4K329 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on 2× RTX 3090 report
2× RTX 40908-bit4K346 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit4K415 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit4K335 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on RTX 6000 Ada report
L40S8-bit4K311 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on L40S report
2× RTX 50908-bit4K489 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on 2× RTX 5090 report
4× RTX 30908-bit4K329 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit4K489 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit4K62 tok/s5.2 sPDF: OLMoE 1B 7B 0924 on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit4K62 tok/s5.2 sPDF: OLMoE 1B 7B 0924 on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit4K62 tok/s5.2 sPDF: OLMoE 1B 7B 0924 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit4K120 tok/s2.6 sPDF: OLMoE 1B 7B 0924 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit4K120 tok/s2.6 sPDF: OLMoE 1B 7B 0924 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit4K120 tok/s2.6 sPDF: OLMoE 1B 7B 0924 on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit4K180 tok/s1.3 sPDF: OLMoE 1B 7B 0924 on M4 Max 128 GB report
M5 Max 128 GB8-bit4K192 tok/s0.4 sPDF: OLMoE 1B 7B 0924 on M5 Max 128 GB report
M3 Ultra 96 GB8-bit4K208 tok/s0.7 sPDF: OLMoE 1B 7B 0924 on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit4K208 tok/s0.7 sPDF: OLMoE 1B 7B 0924 on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit4K208 tok/s0.7 sPDF: OLMoE 1B 7B 0924 on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit4K251 tok/s0.2 sPDF: OLMoE 1B 7B 0924 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit4K97 tok/s0.8 sPDF: OLMoE 1B 7B 0924 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit4K97 tok/s0.8 sPDF: OLMoE 1B 7B 0924 on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit4K134 tok/s0.3 sPDF: OLMoE 1B 7B 0924 on RTX Spark report
DGX Spark8-bit4K124 tok/s0.3 sPDF: OLMoE 1B 7B 0924 on DGX Spark report
Dual DGX Spark8-bit4K124 tok/s0.2 sPDF: OLMoE 1B 7B 0924 on Dual DGX Spark report
DGX Station GB3008-bit4K810 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit4K489 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit4K489 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit4K489 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit4K522 tok/s0.2 sPDF: OLMoE 1B 7B 0924 on A100 80 GB SXM report
H100 SXM8-bit4K649 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on H100 SXM report
4× H200 NVL8-bit4K732 tok/s0.1 sPDF: OLMoE 1B 7B 0924 on 4× H200 NVL report
8× H200 NVL8-bit4K732 tok/s0.1 sPDF: OLMoE 1B 7B 0924 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 allenai/OLMoE-1B-7B-0924-GGUF
vLLM
vllm serve allenai/OLMoE-1B-7B-0924

Questions

How much VRAM does OLMoE 1B 7B 0924 need?

OLMoE 1B 7B 0924 needs 3.9 GB for its weights at 4-bit, 5.0 GB with an 4K context and 5.0 GB with 4K. At 8-bit, 8.0 GB.

What is the minimum VRAM for OLMoE 1B 7B 0924?

About 4.3 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 OLMoE 1B 7B 0924?

6.9B parameters (1.3B active per token): 12.9 GB of weights at 16-bit, 3.9 GB at 4-bit.

Can OLMoE 1B 7B 0924 run on an RTX 4090?

Yes. At 8-bit it holds 4K of context and writes about 346 tokens a second.

More models