readygpu

GLM 4.1V 9B Thinking

10.3B parameters · 64k context · MIT · released 2025-06-28

visionreasoningchat

In short

The cheapest hardware in the US that runs GLM 4.1V 9B Thinking fully in memory at 128k context is the AMD Radeon RX 7900 XT, $800 at Amazon, at about 26 tok/s (estimated).

Updated . Speeds are estimates; how they are worked out.

What to run it on in the US

Fully in memory at 128k context, priced at retailers today.

All hardware that runs it

Everything we price, best value first. Two cards split the model by layers.

HardwareQuantFitsDecodePricetok/s per 1,000
NVIDIA Tesla P40mlx-8bitfits13 tok/s$34536.2
Intel Arc Pro B60mlx-8bitfits13 tok/s$65020.5
AMD Radeon RX 7900 XTmlx-8bitfits26 tok/s$80032.1
Mac mini M4 32GBmlx-8bitfits5 tok/s$1,4203.2
AMD Radeon RX 7900 XTXmlx-8bitfits31 tok/s$1,53320.1
NVIDIA GeForce RTX 3090mlx-8bitfits34 tok/s$1,60021.2
AMD Radeon AI PRO R9700mlx-8bitfits21 tok/s$1,70012.1
NVIDIA GeForce RTX 3090 Timlx-8bitfits37 tok/s$2,06017.7
NVIDIA GeForce RTX 4090mlx-8bitfits37 tok/s$2,07317.6
Mac mini M4 Pro 48GBmlx-8bitfits10 tok/s$2,1494.8
Mac mini M4 Pro 64GBmlx-8bitfits10 tok/s$2,9003.6
Mac Studio M4 Max 64GBmlx-8bitfits21 tok/s$3,4006.1
Framework Desktop Ryzen AI Max+ 395 128GBmlx-8bitfits8 tok/s$3,4492.4
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBmlx-8bitfits8 tok/s$3,6502.2
Mac Studio M3 Ultra 96GBmlx-8bitfits31 tok/s$4,3997.1
NVIDIA GeForce RTX 5090mlx-8bitfits65 tok/s$4,92113.2
Mac Studio M4 Max 128GBmlx-8bitfits21 tok/s$4,9994.1
ASUS Ascent GX10 128GBmlx-8bitfits10 tok/s$6,1001.6
NVIDIA DGX Spark 128GBmlx-8bitfits10 tok/s$7,3991.3
NVIDIA RTX PRO 6000 Blackwellmlx-8bitfits65 tok/s$7,9978.1
Mac Studio M3 Ultra 256GBmlx-8bitfits31 tok/s$8,4993.7
Mac Studio M3 Ultra 512GBmlx-8bitfits31 tok/s$23,0001.4
Framework Desktop Ryzen AI Max+ 395 64GBmlx-8bitfits8 tok/s––
Intel Arc B580mlx-8bitoffload6 tok/s$34616.8
NVIDIA GeForce RTX 3060 12GBmlx-8bitoffload6 tok/s$36016.7
Intel Arc A770 16GBmlx-8bitoffload10 tok/s$40025.3
NVIDIA GeForce RTX 3080 10GBmlx-8bitoffload6 tok/s$46911.9
AMD Radeon RX 7600 XTmlx-8bitoffload7 tok/s$48015.0
AMD Radeon RX 9060 XT 16GBmlx-8bitoffload8 tok/s$52014.8
AMD Radeon RX 7800 XTmlx-8bitoffload11 tok/s$55020.4
NVIDIA GeForce RTX 5060 Ti 16GBmlx-8bitoffload11 tok/s$61817.2
NVIDIA GeForce RTX 4070 Supermlx-8bitoffload7 tok/s$6999.3
NVIDIA GeForce RTX 5070mlx-8bitoffload7 tok/s$7509.2
NVIDIA GeForce RTX 4060 Ti 16GBmlx-8bitoffload8 tok/s$77010.5
AMD Radeon RX 9070 XTmlx-8bitoffload11 tok/s$81514.0
NVIDIA GeForce RTX 4070 Ti Supermlx-8bitoffload13 tok/s$98413.3
NVIDIA GeForce RTX 5070 Timlx-8bitoffload15 tok/s$99515.0
NVIDIA GeForce RTX 4080 Supermlx-8bitoffload14 tok/s$1,19911.4
NVIDIA GeForce RTX 5080mlx-8bitoffload15 tok/s$1,6009.6

1 other card or machine we price cannot hold it at this context.

Memory needed

Weights, KV cache and runtime buffers at each context length (KV cache at fp16).

QuantFile@8k@32k@64k
mlx-8bitmlx11.0 GB12.0 GB13.1 GB14.5 GB

Run it

Copy a command for the runtime you use.

  • mlx-8bit
    • LM Studiolms get mlx-community/GLM-4.1V-9B-Thinking-8bit
    • MLXmlx_lm.server --model mlx-community/GLM-4.1V-9B-Thinking-8bit

Notes

Needs transformers 4.57.1 or later. It is a 9B bilingual (Chinese/English) vision-language reasoning model that handles images up to 4K resolution at any aspect ratio.

Questions

What is the best GPU for running GLM 4.1V 9B Thinking locally?
The cheapest hardware in the US that runs GLM 4.1V 9B Thinking fully in memory at 128k context is the AMD Radeon RX 7900 XT, $800 at Amazon, at about 26 tok/s (estimated). The fastest we price is the NVIDIA GeForce RTX 5090, at about 65 tok/s.
Can I run GLM 4.1V 9B Thinking on a 16 GB GPU?
Only partly: at 128k context it does not fit in 16 GB, so part of it runs from system RAM and it slows to about 10 tok/s on the Intel Arc A770 16GB (estimated).
Can I run GLM 4.1V 9B Thinking on a 24 GB GPU?
Yes: at mlx-8bit it fits fully in 24 GB at 128k context. On the NVIDIA GeForce RTX 3090 Ti it runs at about 37 tok/s (estimated).
What licence is GLM 4.1V 9B Thinking released under?
MIT, which allows commercial use.