---
license: apache-2.0
language: [en, ko, zh]
tags: [qwen3.5, model-merging, dare-ties, darwin-v4, moe, reasoning]
base_model:
- Qwen/Qwen3.5-35B-A3B
- Jackrong/Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled
---

# 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.