KorReason-35B-Darwin

Darwin v4 evolutionary merge: Qwen3.5-35B-A3B Instruct + Claude Opus 4.6 Reasoning Distilled via DARE-TIES.

Merge Config

• Model A: Qwen/Qwen3.5-35B-A3B (Instruct, weight=0.5)
• Model B: Jackrong/Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled (weight=0.5)
• Method: DARE-TIES
• Genome: [global=0.5, attn=0.5, ffn=0.5, embed=0.5, density_a=0.85, density_b=0.85]
• Architecture: qwen3_5_moe (35.95B total, 3B active MoE)
• Size: 71.9GB (bfloat16)

Benchmark (vs Base)

Benchmark	Merged	Base
GPQA Diamond	0.4747	0.4798
MMLU	0.8330	0.8353
ARC-Challenge	0.6600	0.6300

Darwin v4

Produced by the Darwin v4 A2AP evolutionary LLM merging platform.