Korean-preserving one-shot MoE expert pruning of Qwen3-Coder-30B-A3B-Instruct, produced with K-REAP. Routed experts pruned 128 → 112 per layer (12.5%); ~27B 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 78.7 (−0.6pp) with 0/128 collapse, i.e. the Korean loss is reduced to a few points or less. (K-REAP removes ~97% 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	54.9 (−0.3)
KoBEST BoolQ	93.6	92.4 (−1.2)
KoBEST COPA	92.7	91.6 (−1.1)
KoBEST SentiNeg	95.0	94.2 (−0.8)
KoBEST HellaSwag	60.0	60.2 (+0.2)
Gen. collapse (/128, ↓)	0	0

English reasoning/knowledge (llm-evalbox)

benchmark	base	this model
ARC-Challenge	91.2	90.6 (−0.6)
HellaSwag	84.4	86.4 (+2.0)
WinoGrande	73.4	71.0 (−2.4)
MMLU-Pro	50.2	49.8 (−0.4)
TruthfulQA	71.2	72.4 (+1.2)
MMLU-en	76.6	75.0 (−1.6)

Math

benchmark	base	this model
GSM8K	93.2	93.8 (+0.6)
MathQA	55.2	61.2 (+6.0)

Coding

benchmark	base	this model
HumanEval+	87.8	89.6 (+1.8)
MBPP+	77.2	76.2 (−1.0)
LiveCodeBench	52.4	52.9 (+0.5)

Safety/Bias

benchmark	base	this model
BBQ	92.2	91.6 (−0.6)
SafetyBench	83.2	82.2 (−1.0)

Macro

benchmark	base	this model
Korean MC macro	79.3	78.7 (−0.6)
Academic 10-bench macro	74.7	75.2 (+0.5)

Usage (MLX)

python

from mlx_lm import load, generate
model, tok = load("KCh3dRi4n/Qwen3-coder-27B-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