Skip to content

Phi 3.5 MOE Instruct VRAM requirements

How much VRAM Phi 3.5 MOE Instruct needs
4-bit weights: 23.6 GBWith 8K context: 25.3 GBWith 32K context: 28.3 GB8-bit, 8K context: 43.1 GB
Parameters
41.9B
Active per token
6.6B
Longest context
128K
Released
Aug 2024
License
MIT
Strongest at
Not rated

How much VRAM Phi 3.5 MOE Instruct needs

Memory Phi 3.5 MOE Instruct needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit78.0 GB79.6 GB82.6 GB94.6 GB
8-bit41.4 GB43.1 GB46.1 GB58.1 GB
6-bit32.0 GB33.6 GB36.6 GB48.6 GB
5-bit27.7 GB29.4 GB32.4 GB44.4 GB
4-bit23.6 GB25.3 GB28.3 GB40.3 GB
3-bit19.1 GB20.7 GB23.7 GB35.7 GB
2-bit14.6 GB16.3 GB19.3 GB31.3 GB
1.6-bit8.0 GB9.7 GB12.7 GB24.7 GB

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

GPUs and Macs that run Phi 3.5 MOE Instruct

The best way Phi 3.5 MOE Instruct runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Phi 3.5 MOE Instruct on RTX 4060 report
RTX 5060Too bigPDF: Phi 3.5 MOE Instruct on RTX 5060 report
RTX 3070Too bigPDF: Phi 3.5 MOE Instruct on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Phi 3.5 MOE Instruct on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: Phi 3.5 MOE Instruct on RTX 3060 12 GB report
Arc B580Too bigPDF: Phi 3.5 MOE Instruct on Arc B580 report
RTX 4070Too bigPDF: Phi 3.5 MOE Instruct on RTX 4070 report
RTX 4070 SuperToo bigPDF: Phi 3.5 MOE Instruct on RTX 4070 Super report
RTX 5070Too bigPDF: Phi 3.5 MOE Instruct on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: Phi 3.5 MOE Instruct on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: Phi 3.5 MOE Instruct on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: Phi 3.5 MOE Instruct on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: Phi 3.5 MOE Instruct on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: Phi 3.5 MOE Instruct on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: Phi 3.5 MOE Instruct on RTX 4080 Super report
RTX 5070 TiToo bigPDF: Phi 3.5 MOE Instruct on RTX 5070 Ti report
RTX 5080Too bigPDF: Phi 3.5 MOE Instruct on RTX 5080 report
Radeon RX 7900 XTX3-bit24K117 tok/s2.1 sPDF: Phi 3.5 MOE Instruct on Radeon RX 7900 XTX report
RTX 30903-bit24K145 tok/s1.3 sPDF: Phi 3.5 MOE Instruct on RTX 3090 report
RTX 40903-bit24K153 tok/s0.6 sPDF: Phi 3.5 MOE Instruct on RTX 4090 report
Radeon AI PRO R97005-bit16K66 tok/s1.4 sPDF: Phi 3.5 MOE Instruct on Radeon AI PRO R9700 report
RTX 50905-bit16K176 tok/s0.5 sPDF: Phi 3.5 MOE Instruct on RTX 5090 report
2× RTX 30908-bit20K79 tok/s0.8 sPDF: Phi 3.5 MOE Instruct on 2× RTX 3090 report
2× RTX 40908-bit20K84 tok/s0.3 sPDF: Phi 3.5 MOE Instruct on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit28K106 tok/s0.7 sPDF: Phi 3.5 MOE Instruct on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit28K81 tok/s0.5 sPDF: Phi 3.5 MOE Instruct on RTX 6000 Ada report
L40S8-bit28K74 tok/s0.8 sPDF: Phi 3.5 MOE Instruct on L40S report
2× RTX 50908-bit128K133 tok/s0.3 sPDF: Phi 3.5 MOE Instruct on 2× RTX 5090 report
4× RTX 30908-bit128K79 tok/s0.4 sPDF: Phi 3.5 MOE Instruct on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit128K133 tok/s0.4 sPDF: Phi 3.5 MOE Instruct on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Phi 3.5 MOE Instruct on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: Phi 3.5 MOE Instruct on Mac mini M4 24 GB report
Mac mini M4 32 GBToo bigPDF: Phi 3.5 MOE Instruct on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: Phi 3.5 MOE Instruct on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB5-bit48K39 tok/s15 sPDF: Phi 3.5 MOE Instruct on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit28K28 tok/s15 sPDF: Phi 3.5 MOE Instruct on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit128K46 tok/s7.6 sPDF: Phi 3.5 MOE Instruct on M4 Max 128 GB report
M5 Max 128 GB8-bit128K50 tok/s2.2 sPDF: Phi 3.5 MOE Instruct on M5 Max 128 GB report
M3 Ultra 96 GB8-bit128K56 tok/s4.3 sPDF: Phi 3.5 MOE Instruct on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit128K56 tok/s4.3 sPDF: Phi 3.5 MOE Instruct on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit128K56 tok/s4.3 sPDF: Phi 3.5 MOE Instruct on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit128K74 tok/s1.3 sPDF: Phi 3.5 MOE Instruct on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit28K20 tok/s4.4 sPDF: Phi 3.5 MOE Instruct on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit128K20 tok/s4.4 sPDF: Phi 3.5 MOE Instruct on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit128K28 tok/s1.5 sPDF: Phi 3.5 MOE Instruct on RTX Spark report
DGX Spark8-bit128K26 tok/s1.5 sPDF: Phi 3.5 MOE Instruct on DGX Spark report
Dual DGX Spark8-bit128K26 tok/s0.9 sPDF: Phi 3.5 MOE Instruct on Dual DGX Spark report
DGX Station GB3008-bit128K299 tok/s0.1 sPDF: Phi 3.5 MOE Instruct on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit128K133 tok/s0.2 sPDF: Phi 3.5 MOE Instruct on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit128K133 tok/s0.1 sPDF: Phi 3.5 MOE Instruct on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit128K133 tok/s0.1 sPDF: Phi 3.5 MOE Instruct on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit128K146 tok/s1.0 sPDF: Phi 3.5 MOE Instruct on A100 80 GB SXM report
H100 SXM8-bit128K203 tok/s0.3 sPDF: Phi 3.5 MOE Instruct on H100 SXM report
4× H200 NVL8-bit128K249 tok/s0.1 sPDF: Phi 3.5 MOE Instruct on 4× H200 NVL report
8× H200 NVL8-bit128K249 tok/s0.1 sPDF: Phi 3.5 MOE 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 bartowski/Phi-3.5-MoE-instruct-GGUF
vLLM
vllm serve microsoft/Phi-3.5-MoE-instruct

Questions

How much VRAM does Phi 3.5 MOE Instruct need?

Phi 3.5 MOE Instruct needs 23.6 GB for its weights at 4-bit, 25.3 GB with an 8K context and 28.3 GB with 32K. At 8-bit, 43.1 GB.

What is the minimum VRAM for Phi 3.5 MOE Instruct?

About 20.7 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 Radeon RX 7900 XTX.

How big is Phi 3.5 MOE Instruct?

41.9B parameters (6.6B active per token): 78.0 GB of weights at 16-bit, 23.6 GB at 4-bit.

Can Phi 3.5 MOE Instruct run on an RTX 4090?

Yes. At 3-bit it holds 24K of context and writes about 153 tokens a second.

More models