Skip to content

Kimi VL A3B Thinking 2506 VRAM requirements

How much VRAM Kimi VL A3B Thinking 2506 needs
4-bit weights: 9.3 GBWith 8K context: 10.3 GBWith 32K context: 11.0 GB8-bit, 8K context: 17.3 GB
Parameters
16.4B
Active per token
3B
Longest context
128K
Released
Jun 2025
License
MIT
Strongest at
Text and images

How much VRAM Kimi VL A3B Thinking 2506 needs

Memory Kimi VL A3B Thinking 2506 needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit30.6 GB31.6 GB32.3 GB35.2 GB
8-bit16.2 GB17.3 GB18.0 GB20.9 GB
6-bit12.5 GB13.6 GB14.3 GB17.2 GB
5-bit10.9 GB11.9 GB12.6 GB15.5 GB
4-bit9.3 GB10.3 GB11.0 GB13.9 GB
3-bit7.5 GB8.5 GB9.2 GB12.1 GB
2-bit5.7 GB6.8 GB7.5 GB10.4 GB
1.6-bit3.2 GB4.2 GB4.9 GB7.8 GB

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

GPUs and Macs that run Kimi VL A3B Thinking 2506

The best way Kimi VL A3B Thinking 2506 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Kimi VL A3B Thinking 2506 on RTX 4060 report
RTX 5060Too bigPDF: Kimi VL A3B Thinking 2506 on RTX 5060 report
RTX 3070Too bigPDF: Kimi VL A3B Thinking 2506 on RTX 3070 report
RTX 3080 10 GB3-bit40K229 tok/s0.6 sPDF: Kimi VL A3B Thinking 2506 on RTX 3080 10 GB report
RTX 3060 12 GB4-bit44K113 tok/s1.4 sPDF: Kimi VL A3B Thinking 2506 on RTX 3060 12 GB report
Arc B5804-bit44K94 tok/s1.0 sPDF: Kimi VL A3B Thinking 2506 on Arc B580 report
RTX 40704-bit44K148 tok/s0.6 sPDF: Kimi VL A3B Thinking 2506 on RTX 4070 report
RTX 4070 Super4-bit44K148 tok/s0.5 sPDF: Kimi VL A3B Thinking 2506 on RTX 4070 Super report
RTX 50704-bit44K182 tok/s0.6 sPDF: Kimi VL A3B Thinking 2506 on RTX 5070 report
RTX 4060 Ti 16 GB6-bit60K72 tok/s0.8 sPDF: Kimi VL A3B Thinking 2506 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB6-bit60K106 tok/s0.7 sPDF: Kimi VL A3B Thinking 2506 on RTX 5060 Ti 16 GB report
Arc A770 16 GB6-bit60K87 tok/s0.8 sPDF: Kimi VL A3B Thinking 2506 on Arc A770 16 GB report
Radeon RX 9070 XT6-bit60K113 tok/s0.5 sPDF: Kimi VL A3B Thinking 2506 on Radeon RX 9070 XT report
RTX 4070 Ti Super6-bit60K146 tok/s0.4 sPDF: Kimi VL A3B Thinking 2506 on RTX 4070 Ti Super report
RTX 4080 Super6-bit60K156 tok/s0.3 sPDF: Kimi VL A3B Thinking 2506 on RTX 4080 Super report
RTX 5070 Ti6-bit60K180 tok/s0.4 sPDF: Kimi VL A3B Thinking 2506 on RTX 5070 Ti report
RTX 50806-bit60K189 tok/s0.3 sPDF: Kimi VL A3B Thinking 2506 on RTX 5080 report
Radeon RX 7900 XTX8-bit128K126 tok/s0.8 sPDF: Kimi VL A3B Thinking 2506 on Radeon RX 7900 XTX report
RTX 30908-bit128K154 tok/s0.5 sPDF: Kimi VL A3B Thinking 2506 on RTX 3090 report
RTX 40908-bit128K163 tok/s0.2 sPDF: Kimi VL A3B Thinking 2506 on RTX 4090 report
Radeon AI PRO R97008-bit128K93 tok/s0.5 sPDF: Kimi VL A3B Thinking 2506 on Radeon AI PRO R9700 report
RTX 50908-bit128K240 tok/s0.2 sPDF: Kimi VL A3B Thinking 2506 on RTX 5090 report
2× RTX 30908-bit128K154 tok/s0.3 sPDF: Kimi VL A3B Thinking 2506 on 2× RTX 3090 report
2× RTX 40908-bit128K163 tok/s0.1 sPDF: Kimi VL A3B Thinking 2506 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit128K200 tok/s0.3 sPDF: Kimi VL A3B Thinking 2506 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit128K157 tok/s0.2 sPDF: Kimi VL A3B Thinking 2506 on RTX 6000 Ada report
L40S8-bit128K145 tok/s0.3 sPDF: Kimi VL A3B Thinking 2506 on L40S report
2× RTX 50908-bit128K240 tok/s0.1 sPDF: Kimi VL A3B Thinking 2506 on 2× RTX 5090 report
4× RTX 30908-bit128K154 tok/s0.2 sPDF: Kimi VL A3B Thinking 2506 on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit128K240 tok/s0.1 sPDF: Kimi VL A3B Thinking 2506 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB3-bit64K54 tok/s11 sPDF: Kimi VL A3B Thinking 2506 on Mac mini M4 16 GB report
Mac mini M4 24 GB6-bit64K35 tok/s11 sPDF: Kimi VL A3B Thinking 2506 on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit112K28 tok/s11 sPDF: Kimi VL A3B Thinking 2506 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB6-bit64K68 tok/s5.6 sPDF: Kimi VL A3B Thinking 2506 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit128K55 tok/s5.6 sPDF: Kimi VL A3B Thinking 2506 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit128K55 tok/s5.6 sPDF: Kimi VL A3B Thinking 2506 on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit128K86 tok/s2.8 sPDF: Kimi VL A3B Thinking 2506 on M4 Max 128 GB report
M5 Max 128 GB8-bit128K93 tok/s0.8 sPDF: Kimi VL A3B Thinking 2506 on M5 Max 128 GB report
M3 Ultra 96 GB8-bit128K102 tok/s1.6 sPDF: Kimi VL A3B Thinking 2506 on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit128K102 tok/s1.6 sPDF: Kimi VL A3B Thinking 2506 on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit128K102 tok/s1.6 sPDF: Kimi VL A3B Thinking 2506 on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit128K127 tok/s0.5 sPDF: Kimi VL A3B Thinking 2506 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit128K43 tok/s1.6 sPDF: Kimi VL A3B Thinking 2506 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit128K43 tok/s1.6 sPDF: Kimi VL A3B Thinking 2506 on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit128K60 tok/s0.6 sPDF: Kimi VL A3B Thinking 2506 on RTX Spark report
DGX Spark8-bit128K55 tok/s0.6 sPDF: Kimi VL A3B Thinking 2506 on DGX Spark report
Dual DGX Spark8-bit128K55 tok/s0.3 sPDF: Kimi VL A3B Thinking 2506 on Dual DGX Spark report
DGX Station GB3008-bit128K442 tok/s0.1 sPDF: Kimi VL A3B Thinking 2506 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit128K240 tok/s0.1 sPDF: Kimi VL A3B Thinking 2506 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit128K240 tok/s0.1 sPDF: Kimi VL A3B Thinking 2506 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit128K240 tok/s0.1 sPDF: Kimi VL A3B Thinking 2506 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit128K260 tok/s0.4 sPDF: Kimi VL A3B Thinking 2506 on A100 80 GB SXM report
H100 SXM8-bit128K336 tok/s0.1 sPDF: Kimi VL A3B Thinking 2506 on H100 SXM report
4× H200 NVL8-bit128K389 tok/s0.1 sPDF: Kimi VL A3B Thinking 2506 on 4× H200 NVL report
8× H200 NVL8-bit128K389 tok/s0.1 sPDF: Kimi VL A3B Thinking 2506 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/Kimi-VL-A3B-Thinking-2506-GGUF
vLLM
vllm serve moonshotai/Kimi-VL-A3B-Thinking-2506

Questions

How much VRAM does Kimi VL A3B Thinking 2506 need?

Kimi VL A3B Thinking 2506 needs 9.3 GB for its weights at 4-bit, 10.3 GB with an 8K context and 11.0 GB with 32K. At 8-bit, 17.3 GB.

What is the minimum VRAM for Kimi VL A3B Thinking 2506?

About 8.5 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 Kimi VL A3B Thinking 2506?

16.4B parameters (3B active per token): 30.6 GB of weights at 16-bit, 9.3 GB at 4-bit.

Can Kimi VL A3B Thinking 2506 run on an RTX 4090?

Yes. At 8-bit it holds 128K of context and writes about 163 tokens a second.

More models