Feb 4: Qwen3-Coder-Next GGUFs reuploaded - much better outputs! (Still in progress)
llama.cpp has fixed a bug which caused the model to loop and produce poor outputs. The calculation for vectorized key_gdiff has been corrected.
Thanks to the work of llama.cpp and contributors, we have now have reconverted and re-uploaded the model.
Please re-download and update llama.cpp thanks!
A lot have been updated, the rest will reupload in an hour or so.
See file history for last updated ones.
Please let us know if you see an improvement!
Q8, MXFP4, F16 are not updated.
We also made a new tutorial on running our dynamic FP8 quant and have a new MXFP4 GGUF.
Guide: https://unsloth.ai/docs/models/qwen3-coder-next
Neither MXFP4 nor Q8 variants have been updated, is this intended or should we expect an update for those quants as well? Thanks for your hard work!
Looks marvelous...
Any plans to roll out a REAP version?
Your https://huggingface.co/unsloth/GLM-4.7-Flash-REAP-23B-A3B-GGUF has fantastic results.
I'm getting a lot of '"filePath"/home/' invalid json syntax in tool calls (with Q6_K_XL in opencode), and looping instead of fixing it (even when told to fix it, annoyingly). Is this why? Will pulling the files down again fix that?
Neither MXFP4 nor Q8 variants have been updated, is this intended or should we expect an update for those quants as well? Thanks for your hard work!
Those quants do not use imatrix, so they are fine to be used as-is. Only quants using imatrixes needed to be requantized.
Those quants do not use imatrix, so they are fine to be used as-is. Only quants using imatrixes needed to be requantized.
But Q2-Q6 are re-uploaded too. They don't use imatrix I think. I don't understand.
Neither MXFP4 nor Q8 variants have been updated, is this intended or should we expect an update for those quants as well? Thanks for your hard work!
Those are not imatrix so it's not needed
Those quants do not use imatrix, so they are fine to be used as-is. Only quants using imatrixes needed to be requantized.
But Q2-Q6 are re-uploaded too. They don't use imatrix I think. I don't understand.
They are imatrix. Only ones arent are 8bit and above and MXFP$