Skip to content

Moonlight 16B A3B Instruct VRAM requirements

How much VRAM Moonlight 16B A3B Instruct needs
4-bit weights: 9.0 GBWith 8K context: 10.1 GB8-bit, 8K context: 16.9 GB
Parameters
16B
Active per token
3B
Longest context
8K
Released
Feb 2025
License
MIT
Strongest at
Not rated

How much VRAM Moonlight 16B A3B Instruct needs

Memory Moonlight 16B A3B Instruct needs at each precision and context length
PrecisionWeightsWith 8K
16-bit29.7 GB30.8 GB
8-bit15.8 GB16.9 GB
6-bit12.2 GB13.3 GB
5-bit10.6 GB11.6 GB
4-bit9.0 GB10.1 GB
3-bit7.3 GB8.3 GB
2-bit5.6 GB6.6 GB
1.6-bit3.1 GB4.1 GB

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

GPUs and Macs that run Moonlight 16B A3B Instruct

The best way Moonlight 16B A3B Instruct runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Moonlight 16B A3B Instruct on RTX 4060 report
RTX 5060Too bigPDF: Moonlight 16B A3B Instruct on RTX 5060 report
RTX 3070Too bigPDF: Moonlight 16B A3B Instruct on RTX 3070 report
RTX 3080 10 GB3-bit8K229 tok/s0.6 sPDF: Moonlight 16B A3B Instruct on RTX 3080 10 GB report
RTX 3060 12 GB5-bit8K99 tok/s1.4 sPDF: Moonlight 16B A3B Instruct on RTX 3060 12 GB report
Arc B5805-bit8K83 tok/s1.0 sPDF: Moonlight 16B A3B Instruct on Arc B580 report
RTX 40705-bit8K130 tok/s0.6 sPDF: Moonlight 16B A3B Instruct on RTX 4070 report
RTX 4070 Super5-bit8K130 tok/s0.5 sPDF: Moonlight 16B A3B Instruct on RTX 4070 Super report
RTX 50705-bit8K162 tok/s0.6 sPDF: Moonlight 16B A3B Instruct on RTX 5070 report
RTX 4060 Ti 16 GB6-bit8K72 tok/s0.8 sPDF: Moonlight 16B A3B Instruct on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB6-bit8K106 tok/s0.7 sPDF: Moonlight 16B A3B Instruct on RTX 5060 Ti 16 GB report
Arc A770 16 GB6-bit8K87 tok/s0.8 sPDF: Moonlight 16B A3B Instruct on Arc A770 16 GB report
Radeon RX 9070 XT6-bit8K113 tok/s0.5 sPDF: Moonlight 16B A3B Instruct on Radeon RX 9070 XT report
RTX 4070 Ti Super6-bit8K146 tok/s0.4 sPDF: Moonlight 16B A3B Instruct on RTX 4070 Ti Super report
RTX 4080 Super6-bit8K156 tok/s0.3 sPDF: Moonlight 16B A3B Instruct on RTX 4080 Super report
RTX 5070 Ti6-bit8K180 tok/s0.4 sPDF: Moonlight 16B A3B Instruct on RTX 5070 Ti report
RTX 50806-bit8K189 tok/s0.3 sPDF: Moonlight 16B A3B Instruct on RTX 5080 report
Radeon RX 7900 XTX8-bit8K126 tok/s0.8 sPDF: Moonlight 16B A3B Instruct on Radeon RX 7900 XTX report
RTX 30908-bit8K154 tok/s0.5 sPDF: Moonlight 16B A3B Instruct on RTX 3090 report
RTX 40908-bit8K163 tok/s0.2 sPDF: Moonlight 16B A3B Instruct on RTX 4090 report
Radeon AI PRO R97008-bit8K93 tok/s0.5 sPDF: Moonlight 16B A3B Instruct on Radeon AI PRO R9700 report
RTX 50908-bit8K240 tok/s0.2 sPDF: Moonlight 16B A3B Instruct on RTX 5090 report
2× RTX 30908-bit8K154 tok/s0.3 sPDF: Moonlight 16B A3B Instruct on 2× RTX 3090 report
2× RTX 40908-bit8K163 tok/s0.1 sPDF: Moonlight 16B A3B Instruct on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit8K200 tok/s0.3 sPDF: Moonlight 16B A3B Instruct on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit8K157 tok/s0.2 sPDF: Moonlight 16B A3B Instruct on RTX 6000 Ada report
L40S8-bit8K145 tok/s0.3 sPDF: Moonlight 16B A3B Instruct on L40S report
2× RTX 50908-bit8K240 tok/s0.1 sPDF: Moonlight 16B A3B Instruct on 2× RTX 5090 report
4× RTX 30908-bit8K154 tok/s0.2 sPDF: Moonlight 16B A3B Instruct on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit8K240 tok/s0.1 sPDF: Moonlight 16B A3B Instruct on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB4-bit8K45 tok/s11 sPDF: Moonlight 16B A3B Instruct on Mac mini M4 16 GB report
Mac mini M4 24 GB6-bit8K35 tok/s11 sPDF: Moonlight 16B A3B Instruct on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit8K28 tok/s11 sPDF: Moonlight 16B A3B Instruct on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB6-bit8K68 tok/s5.6 sPDF: Moonlight 16B A3B Instruct on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit8K55 tok/s5.6 sPDF: Moonlight 16B A3B Instruct on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit8K55 tok/s5.6 sPDF: Moonlight 16B A3B Instruct on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit8K86 tok/s2.8 sPDF: Moonlight 16B A3B Instruct on M4 Max 128 GB report
M5 Max 128 GB8-bit8K93 tok/s0.8 sPDF: Moonlight 16B A3B Instruct on M5 Max 128 GB report
M3 Ultra 96 GB8-bit8K102 tok/s1.6 sPDF: Moonlight 16B A3B Instruct on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit8K102 tok/s1.6 sPDF: Moonlight 16B A3B Instruct on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit8K102 tok/s1.6 sPDF: Moonlight 16B A3B Instruct on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit8K127 tok/s0.5 sPDF: Moonlight 16B A3B Instruct on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit8K43 tok/s1.6 sPDF: Moonlight 16B A3B Instruct on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit8K43 tok/s1.6 sPDF: Moonlight 16B A3B Instruct on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit8K60 tok/s0.6 sPDF: Moonlight 16B A3B Instruct on RTX Spark report
DGX Spark8-bit8K55 tok/s0.6 sPDF: Moonlight 16B A3B Instruct on DGX Spark report
Dual DGX Spark8-bit8K55 tok/s0.3 sPDF: Moonlight 16B A3B Instruct on Dual DGX Spark report
DGX Station GB3008-bit8K442 tok/s0.1 sPDF: Moonlight 16B A3B Instruct on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit8K240 tok/s0.1 sPDF: Moonlight 16B A3B Instruct on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit8K240 tok/s0.1 sPDF: Moonlight 16B A3B Instruct on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit8K240 tok/s0.1 sPDF: Moonlight 16B A3B Instruct on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit8K260 tok/s0.4 sPDF: Moonlight 16B A3B Instruct on A100 80 GB SXM report
H100 SXM8-bit8K336 tok/s0.1 sPDF: Moonlight 16B A3B Instruct on H100 SXM report
4× H200 NVL8-bit8K389 tok/s0.1 sPDF: Moonlight 16B A3B Instruct on 4× H200 NVL report
8× H200 NVL8-bit8K389 tok/s0.1 sPDF: Moonlight 16B A3B 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 mmnga/Moonlight-16B-A3B-Instruct-gguf
vLLM
vllm serve moonshotai/Moonlight-16B-A3B-Instruct
MLX (Mac)
mlx_lm.chat --model mlx-community/Moonlight-16B-A3B-Instruct-4-bit

Questions

How much VRAM does Moonlight 16B A3B Instruct need?

Moonlight 16B A3B Instruct needs 9.0 GB for its weights at 4-bit, 10.1 GB with an 8K context and 10.1 GB with 8K. At 8-bit, 16.9 GB.

What is the minimum VRAM for Moonlight 16B A3B Instruct?

About 8.3 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 3080 10 GB.

How big is Moonlight 16B A3B Instruct?

16B parameters (3B active per token): 29.7 GB of weights at 16-bit, 9.0 GB at 4-bit.

Can Moonlight 16B A3B Instruct run on an RTX 4090?

Yes. At 8-bit it holds 8K of context and writes about 163 tokens a second.

More models