Modelpublic

anik-jha/Qwen3.6-35B-A3B-coding-reap50-GGUF

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

7likes1.1kdownloads

Model Card

Qwen3.6-35B-A3B coding specialist, 50% experts pruned (GGUF)

Qwen3.6-35B-A3B with half of its experts removed (128 of 256 per layer, REAP scoring on a coding-heavy calibration mix), quantized with an importance matrix. This is the main artifact of the paper Half the Experts, All the Code (arXiv link coming): a 19B-parameter coding specialist that fits where the full model does not.

Why you might want it: at the same memory budget, these beat 2-bit quantization of the full model on HumanEval+.

file	size	HumanEval+	MBPP+
Q4KM (imatrix)	11.4 GB	0.902	0.720
Q5KM (imatrix)	13.3 GB	0.915	0.749
full model, IQ2_M (for comparison, not included)	13.0 GB	0.896	0.730
full model, Q8 (for comparison, not included)	38 GB	0.890	0.772

All numbers are greedy pass@1 via EvalPlus against llama.cpp, reasoning off. The trade is real: general (non-coding) perplexity roughly doubles. This is a coding specialist, not a general assistant - don't deploy it as one.

Made with moep; the exact expert selection JSON is in that repo under artifacts/selections/, so you can reproduce the surgery from the base checkpoint. Runs anywhere llama.cpp runs.

Derivative of Qwen3.6-35B-A3B, Apache-2.0, upstream notice retained.