Qwen3.6-35B-A3B coding specialist, 25% experts kept, healed (GGUF)

Companion checkpoint for Half the Experts, All the Code: One-Shot Domain Pruning of Mixture-of-Experts LLMs for Coding (arXiv:2607.16721), accepted at the COLM 2026 Workshop on Efficient Reasoning (OpenReview). Pipeline and the exact keep-set JSON that produced it: github.com/anik-jha/moep.

75% of the routed experts were removed from Qwen3.6-35B-A3B under the REAP criterion calibrated on a coding-weighted mix, then the survivors were healed by distilling from the base model with the expert tensors frozen (LoRA on attention, router, and shared expert only).

Read this before using it

This is a research artifact, and the paper's headline finding is a negative one: at this memory budget you are better off quantizing the full model than pruning it. The healed checkpoint below sits at 11.2 GB and scores 0.841 on HumanEval+. A 2-bit IQ2_M quantization of the unpruned model sits at 12.1 GB and scores 0.896. Pruning to 25% and healing does not close that gap; five separate attempts to close it are reported in the paper and all failed.

The checkpoint is released because healing is GPU-nondeterministic and therefore not reproducible from the keep-set JSON, unlike every other point in the study. It is the artifact behind the paper's healing result, not a recommended deployment. If you want something to actually run, use the 50%-keep specialist: Qwen3.6-35B-A3B-coding-reap50-GGUF.

Evaluations

All numbers from runs/qwen36/eval.json in the pipeline repo, greedy decoding, EvalPlus HumanEval+ and MBPP+ (pass@1). Only the Q8_0 file was functionally evaluated.

Healing recovers 54 of the 103 HumanEval+ points lost to pruning (0.787 → 0.841 against a base of 0.890) and 26 of 113 on MBPP+. Held-out perplexity improves from 5.45 to 4.85 on code and from 30.9 to 9.4 on general text.

The Q4_K_M file in this repo has not been functionally evaluated. It passed an export smoke test and was benchmarked for throughput, nothing more. Do not attribute the table above to it.

Files

Throughput measured with llama-bench on one DGX Spark (GB10, 119 GB unified memory): Q8_0 60.7 tok/s generation, 2780 tok/s prefill; Q4_K_M 82.0 tok/s generation, 3081 tok/s prefill.

License

Weights are a derivative of Qwen3.6-35B-A3B and inherit its Apache-2.0 license. The pipeline that produced them is MIT.

Citation

@misc{jha2026halfexperts,
  title         = {Half the Experts, All the Code: One-Shot Domain Pruning of Mixture-of-Experts LLMs for Coding},
  author        = {Jha, Anik},
  year          = {2026},
  eprint        = {2607.16721},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  note          = {Accepted at the COLM 2026 Workshop on Efficient Reasoning},
  url           = {https://arxiv.org/abs/2607.16721}
}