Memory needed vs common local setups (UD-Q4_K_XL 249.4 GB)
Thanks for the quants. For anyone checking whether this fits their machine, quick math from the file sizes in this repo (HF API, 23 Oct):
- UD-Q4_K_XL: 249.4 GB of files (7 shards)
- UD-Q8_K_XL: 259.1 GB (only ~4% bigger, since the experts already ship in 4-bit)
That's files only, before KV cache and runtime overhead. Usable memory on common setups (our estimates):
- 24 GB GPU: ~23.4 GB
- 32 GB GPU: ~31.4 GB
- 128 GB unified (Mac / DGX Spark class): ~96 GB
- 256 GB Mac Studio: ~192 GB
So UD-Q4_K_XL doesn't fit fully on any single consumer box; you need about 256 GB+ of GPU memory, or offload most experts to system RAM (with 23B active per token, CPU offload is usable but slow). A 2-bit quant (about 156 GB in the community repos) is the only size that fits under ~192 GB.
We keep per-GPU fit estimates here: https://localllmrig.com/can-i-run?utm_source=huggingface&utm_medium=community&utm_campaign=minimax-m3-1-flash
Happy to correct anything if someone has measured numbers.
This model require too much resources and very slow compared to Deepseek V4 1017, i'll pass, its unusable. Launching it with reasoning off not working at all, its mandatory here, with low reasoning it failed simple music notes test.
I used the Q8_K_XLARGE (only high quality, no compromises) and compared to Deepseek Q8_K_XLARGE by Unsloth also the problem started already in allocation resources, Minimax uses much more GPU memory to even launch, on my RTX 4070 16Gb (13 layers vs 22 layers in Deepseek).
And generation speed almost x2 slower in Minimax vs Deeepseek by tok/sec.
Yes Deepseek use many tricks to speed up like Engrams and etc.
With Deepseek i can merge many llama git versions overnight when -50% energy discount and no mistakes, Kalman filtering plugin helping and high reasoning level not affecting speed all in Q8 (300Gb RAM). I cant imagine such in Minimax, they need to make something to speed up model. Only if MXFP4 GGUF will be made maybe i'll check Minimax again.
Thanks, measured reports like this are exactly what file sizes can't tell you. Useful datapoint: on an RTX 4070 16 GB at Q8_K_XL, MiniMax fit 13 layers on the GPU vs 22 for DeepSeek V4.1 Flash 1017, at roughly half the generation speed.
If you have them handy, could you share the tok/s numbers (prompt and generation), the context size and the llama.cpp build for both runs? With your OK we'd add them as a credited user-reported row next to our estimates.