Hi Patrick, tried the Q2_K with a prompt "Write a C# 6.0 RPG like Skyrim, no graphics only backend"
It started outputting gibberish in the Think tokens, random symbols like *#$ and words/sentences that made no sense
Temp 1.0, top-p 0.95 as Xiaomi recommended on latest llama-cpp ROCm
Hello, thanks for letting me know! I will do more evaluation this week and see which quantizations and formats work well.
I'm using a 128gb unified ram device, but you might be able to point claude or codex at these repos and the base unpruned teacher Mimo-v2.6-Flash, then prune that model 50% with a calibration set more weighted to game development. Its an option that make a model better suited towards game-oriented fine tuning.
The one good recent REAP of similar size that I saw was "https://huggingface.co/AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF". This was a second attempt, because his first attempt shaved off too many experts and was unstable. Perhaps 50% is too much for MiMo V2.6 ?
One thing nice about MiMo vs your GLM 5.3 Flash prune is that I could launch it out of the box with llama.cpp. I have a unified memory rig which is slow, and a fast dedicated GPU cluster with driver issues (MI50). I spent days rebuilding your GLM 5.3 fork for Vulkan and ROCm.
Vulkan on dedicated GPUs was getting me < 0.1 tok/s, Unified memory < 3 tok/s. After a lot of debugging, I just got my ROCm setup working ~ 7tok/s and am beginning to run longer tests on your GLM 5.3 quants. So far they seem really good.
