Skip to content

Pixtral 12B VRAM requirements

How much VRAM Pixtral 12B needs
4-bit weights: 7.2 GBWith 8K context: 9.3 GBWith 32K context: 13.0 GB8-bit, 8K context: 14.7 GB
Parameters
12.7B
Longest context
1000K
Released
Sep 2024
License
Apache 2.0
Strongest at
Text and images

How much VRAM Pixtral 12B needs

Memory Pixtral 12B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit23.6 GB25.7 GB29.5 GB44.5 GB
8-bit12.5 GB14.7 GB18.4 GB33.4 GB
6-bit9.7 GB11.8 GB15.5 GB30.5 GB
5-bit8.4 GB10.5 GB14.3 GB29.3 GB
4-bit7.2 GB9.3 GB13.0 GB28.0 GB
3-bit5.8 GB7.9 GB11.6 GB26.6 GB
2-bit4.4 GB6.5 GB10.3 GB25.3 GB
1.6-bit2.4 GB4.5 GB8.3 GB23.3 GB

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

GPUs and Macs that run Pixtral 12B

The best way Pixtral 12B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Pixtral 12B on RTX 4060 report
RTX 5060Too bigPDF: Pixtral 12B on RTX 5060 report
RTX 3070Too bigPDF: Pixtral 12B on RTX 3070 report
RTX 3080 10 GB3-bit16K82 tok/s0.9 sPDF: Pixtral 12B on RTX 3080 10 GB report
RTX 3060 12 GB4-bit20K33 tok/s2.0 sPDF: Pixtral 12B on RTX 3060 12 GB report
Arc B5804-bit20K29 tok/s1.4 sPDF: Pixtral 12B on Arc B580 report
RTX 40704-bit20K46 tok/s0.9 sPDF: Pixtral 12B on RTX 4070 report
RTX 4070 Super4-bit20K46 tok/s0.7 sPDF: Pixtral 12B on RTX 4070 Super report
RTX 50704-bit20K60 tok/s0.8 sPDF: Pixtral 12B on RTX 5070 report
RTX 4060 Ti 16 GB6-bit28K20 tok/s1.1 sPDF: Pixtral 12B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB6-bit28K31 tok/s1.1 sPDF: Pixtral 12B on RTX 5060 Ti 16 GB report
Arc A770 16 GB6-bit28K27 tok/s1.2 sPDF: Pixtral 12B on Arc A770 16 GB report
Radeon RX 9070 XT6-bit28K36 tok/s0.7 sPDF: Pixtral 12B on Radeon RX 9070 XT report
RTX 4070 Ti Super6-bit28K45 tok/s0.6 sPDF: Pixtral 12B on RTX 4070 Ti Super report
RTX 4080 Super6-bit28K49 tok/s0.5 sPDF: Pixtral 12B on RTX 4080 Super report
RTX 5070 Ti6-bit28K59 tok/s0.6 sPDF: Pixtral 12B on RTX 5070 Ti report
RTX 50806-bit28K63 tok/s0.5 sPDF: Pixtral 12B on RTX 5080 report
Radeon RX 7900 XTX8-bit56K41 tok/s1.1 sPDF: Pixtral 12B on Radeon RX 7900 XTX report
RTX 30908-bit56K48 tok/s0.7 sPDF: Pixtral 12B on RTX 3090 report
RTX 40908-bit56K52 tok/s0.3 sPDF: Pixtral 12B on RTX 4090 report
Radeon AI PRO R97008-bit104K28 tok/s0.7 sPDF: Pixtral 12B on Radeon AI PRO R9700 report
RTX 50908-bit104K88 tok/s0.2 sPDF: Pixtral 12B on RTX 5090 report
2× RTX 30908-bit200K48 tok/s0.4 sPDF: Pixtral 12B on 2× RTX 3090 report
2× RTX 40908-bit200K52 tok/s0.2 sPDF: Pixtral 12B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit200K68 tok/s0.4 sPDF: Pixtral 12B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit200K50 tok/s0.3 sPDF: Pixtral 12B on RTX 6000 Ada report
L40S8-bit200K45 tok/s0.4 sPDF: Pixtral 12B on L40S report
2× RTX 50908-bit288K88 tok/s0.1 sPDF: Pixtral 12B on 2× RTX 5090 report
4× RTX 30908-bit480K48 tok/s0.2 sPDF: Pixtral 12B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit480K88 tok/s0.2 sPDF: Pixtral 12B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB3-bit20K16 tok/s16 sPDF: Pixtral 12B on Mac mini M4 16 GB report
Mac mini M4 24 GB6-bit28K9.7 tok/s16 sPDF: Pixtral 12B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit44K7.5 tok/s16 sPDF: Pixtral 12B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB6-bit28K21 tok/s8.1 sPDF: Pixtral 12B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit128K17 tok/s8.1 sPDF: Pixtral 12B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit200K17 tok/s8.1 sPDF: Pixtral 12B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit480K29 tok/s4.0 sPDF: Pixtral 12B on M4 Max 128 GB report
M5 Max 128 GB8-bit480K32 tok/s1.2 sPDF: Pixtral 12B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit352K37 tok/s2.3 sPDF: Pixtral 12B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit1000K37 tok/s2.3 sPDF: Pixtral 12B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit1000K37 tok/s2.3 sPDF: Pixtral 12B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit1000K51 tok/s0.7 sPDF: Pixtral 12B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit200K12 tok/s2.3 sPDF: Pixtral 12B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit480K12 tok/s2.3 sPDF: Pixtral 12B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit640K16 tok/s0.8 sPDF: Pixtral 12B on RTX Spark report
DGX Spark8-bit640K15 tok/s0.8 sPDF: Pixtral 12B on DGX Spark report
Dual DGX Spark8-bit1000K15 tok/s0.5 sPDF: Pixtral 12B on Dual DGX Spark report
DGX Station GB3008-bit1000K265 tok/s0.1 sPDF: Pixtral 12B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit1000K88 tok/s0.1 sPDF: Pixtral 12B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit1000K88 tok/s0.1 sPDF: Pixtral 12B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit1000K88 tok/s0.1 sPDF: Pixtral 12B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit384K99 tok/s0.5 sPDF: Pixtral 12B on A100 80 GB SXM report
H100 SXM8-bit384K151 tok/s0.2 sPDF: Pixtral 12B on H100 SXM report
4× H200 NVL8-bit1000K200 tok/s0.1 sPDF: Pixtral 12B on 4× H200 NVL report
8× H200 NVL8-bit1000K200 tok/s0.1 sPDF: Pixtral 12B 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/pixtral-12b-GGUF
vLLM
vllm serve mistral-experimental/pixtral-12b

Questions

How much VRAM does Pixtral 12B need?

Pixtral 12B needs 7.2 GB for its weights at 4-bit, 9.3 GB with an 8K context and 13.0 GB with 32K. At 8-bit, 14.7 GB.

What is the minimum VRAM for Pixtral 12B?

About 7.9 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 Pixtral 12B?

12.7B parameters: 23.6 GB of weights at 16-bit, 7.2 GB at 4-bit.

Can Pixtral 12B run on an RTX 4090?

Yes. At 8-bit it holds 56K of context and writes about 52 tokens a second.

More models