Qwen3-Coder-23B-A3B-KREAP-MLX (25% routed pruning)
Korean-preserving one-shot MoE expert pruning of Qwen3-Coder-30B-A3B-Instruct, produced with K-REAP. Routed experts pruned 128 → 96 per layer (25%); ~23B total params, 3.3B active (top-8 unchanged). No fine-tuning.
Unlike vanilla REAP (English/code calibration), K-REAP detects and hard-protects the experts that carry Korean, so pruning preserves Korean without sacrificing English/coding. See K-Guard-REAP for the full study.
Korean loss minimized vs. standard REAP
The standard REAP recipe collapses Korean at comparable compression: Cerebras Qwen3-Coder-REAP-25B (19.5%, no language protection) drops Korean MC macro 79.3 → 56.6 (−22.7pp) with generation collapse 115/128. By hard-protecting the Korean expert path, this model keeps Korean MC macro at 77.4 (−1.9pp) with 0/128 collapse, i.e. the Korean loss is reduced to a few points or less. (K-REAP removes ~92% of REAP's Korean damage.)
Benchmarks (vs base Qwen3-Coder-30B; Δ = base delta, %p)
한국어 (Korean, self-harness)
| benchmark | base | this model |
|---|
| KMMLU | 55.2 | 51.8 (−3.4) |
| KoBEST BoolQ | 93.6 | 91.0 (−2.6) |
| KoBEST COPA | 92.7 | 88.3 (−4.4) |
| KoBEST SentiNeg | 95.0 | 93.7 (−1.3) |
| KoBEST HellaSwag | 60.0 | 62.0 (+2.0) |
| Gen. collapse (/128, ↓) | 0 | 0 |
English reasoning/knowledge (llm-evalbox)
| benchmark | base | this model |
|---|
| ARC-Challenge | 91.2 | 91.0 (−0.2) |
| HellaSwag | 84.4 | 81.0 (−3.4) |
| WinoGrande | 73.4 | 64.8 (−8.6) |
| MMLU-Pro | 50.2 | 46.8 (−3.4) |
| TruthfulQA | 71.2 | 69.2 (−2.0) |
| MMLU-en | 76.6 | 70.4 (−6.2) |
Math
| benchmark | base | this model |
|---|
| GSM8K | 93.2 | 93.6 (+0.4) |
| MathQA | 55.2 | 63.0 (+7.8) |
Coding
| benchmark | base | this model |
|---|
| HumanEval+ | 87.8 | 87.2 (−0.6) |
| MBPP+ | 77.2 | 75.1 (−2.1) |
| LiveCodeBench | 52.4 | 50.7 (−1.7) |
Safety/Bias
| benchmark | base | this model |
|---|
| BBQ | 92.2 | 92.6 (+0.4) |
| SafetyBench | 83.2 | 80.4 (−2.8) |
Macro
| benchmark | base | this model |
|---|
| Korean MC macro | 79.3 | 77.4 (−1.9) |
| Academic 10-bench macro | 74.7 | 73.3 (−1.4) |
Usage (MLX)
from mlx_lm import load, generate
model, tok = load("RAS-LAB/Qwen3-Coder-23B-A3B-KREAP-MLX")
print(generate(model, tok, prompt="한국어로 자기소개를 해줘.", max_tokens=256))
Standard HF safetensors (BF16) — also loadable with transformers. Coverage note: KoBEST & EvalPlus are full sets; KMMLU/evalbox/LiveCodeBench are deterministic subsets (seed 42). Effect sizes ≫ sampling CI. Details: K-Guard-REAP docs/10_MASTER_RESULTS.md §7.
Method
REAP saliency (router gate × expert-output L2 norm, conditional mean) restricted to Korean segments + Korean↔English contrast + rare/rollout protection → hard-protected survivor set → streaming safetensors surgery (uniform experts/layer). Framework: https://github.com/Chedrian07/K-REAP