Please support my work: https://donate.sybilsolutions.ai

Qwen3.6-28B-REAP20-Opus-A3B

A 20%-expert-pruned + Opus-trace fine-tuned variant of Qwen/Qwen3.6-35B-A3B, produced via Cerebras REAP (Router-weighted Expert Activation Pruning, arXiv:2510.13999) followed by LoRA SFT on public Claude Opus reasoning traces.

Headline numbers

Metric	Base Qwen3.6-35B-A3B	This model (20% REAP + Opus SFT)	Δ
MMLU (200-sample lite)	{{MMLU_BASE}}	{{MMLU_THIS}}	{{MMLU_DELTA}}
GSM8K (100-sample lite)	{{GSM_BASE}}	{{GSM_THIS}}	{{GSM_DELTA}}
HumanEval (50 parse-rate)	{{HE_BASE}}	{{HE_THIS}}	{{HE_DELTA}}
Structured JSON parse (20)	{{JSON_BASE}}	{{JSON_THIS}}	{{JSON_DELTA}}
Mermaid render (10)	{{MERM_BASE}}	{{MERM_THIS}}	{{MERM_DELTA}}
AdvBench refusal (32)	{{REFUSE_BASE}}	{{REFUSE_THIS}}	{{REFUSE_DELTA}}

Architecture

• Base: Qwen3.6-35B-A3B (40 layers, 256 experts/layer, 8 routed + 1 shared active, qwen3_5_moe)
• After 20% REAP: 205 experts/layer kept, 51 experts/layer pruned → ~28B total params, still ~3B active
• Fine-tune: LoRA rank 32, α 64 on q,k,v,o,gate,up,down projections. bf16 weights after merge.

Pipeline

1. Calibration merge — 5,000 stratified samples from:
   • /Users/sero/.../reap-expert-swap/dataset/calibration-20k.jsonl (general, coding, reasoning, etc.)
   • 0xSero/structured-outputs-calibration-v1 (JSON / Mermaid / schema)
2. REAP observation (this fork's Qwen3_5Moe-aware observer, multi-GPU layerwise on 8× A100-40GB): {{OBS_DURATION}}
3. REAP prune @ 20% using reap saliency metric, renormalized router weights, seed 42.
4. Opus-trace SFT via LLaMA-Factory + DeepSpeed ZeRO-3 (8× A100). LoRA 2 epochs on nohurry/Opus-4.6-Reasoning-3000x-filtered (2,326 reasoning trajectories with explicit <think>…</think>\nanswer structure).
5. GGUF — bf16, Q8_0, Q6_K, Q5_K_M, Q4_K_M with imatrix from merged calibration.

Sidecar observations

REAP observation artifacts live in the separate dataset repo 0xSero/qwen3.6-35b-a3b-reap-observations.

Known limitations

• Refusal behavior follows the base model plus Opus SFT; no explicit abliteration was applied in this release. The model will refuse straight adversarial probes at roughly base-model rates.
• Reasoning quality on GSM8K-style problems depends on the <think> chain-of-thought; short max-tokens limits hurt accuracy.
• Structured-output calibration is oversampled vs. base mix (JSON/Mermaid experts preferentially retained).

License

Apache 2.0, inherited from base model. This checkpoint is a derivative work; please preserve attribution.

Citation

@misc{lasby2025reap,
  title   = {REAP: Router-weighted Expert Activation Pruning for Mixture-of-Experts},
  author  = {Lasby, Mike and others},
  year    = {2025},
  eprint  = {2510.13999},
  archivePrefix = {arXiv},
}