REAP 3bit MLX
I'm interested to know if it's possible to convert this into a 3 bit MLX with the REAP Applied like was done here on GLM 7.4
https://huggingface.co/scaryrawr/GLM-4.7-REAP-50-mlx-3Bit
They've managed to get this model weights down to 80 gigabytes. I've got a 96 gigabyte M2 Mac and was wondering if I could do the same here with KIMI K2.5
Feasible -- yes, though REAP pruning beyond 25% hasn't yielded results I would put my name on or stand behind to release for public open use. The full PRISM model remains too large and too expensive for general distribution (I used 16 x B300 for over 100 hours just to work on it for initial PRISM fine-tuning), PRISM tensor access for Kimi-K2.5-PRISM is vaulted for enterprise sponsor/customer usage. If interested in commissioning custom licensed work: https://ko-fi.com/s/a996811f5d