Darwin v4 — Qwen3.5-35B-A3B DARE-TIES Merge
Merged model: 60% Instruct + 40% Base using DARE-TIES algorithm.
Merge Config
- Model A: (Instruct, weight=0.6)
- Model B: (weight=0.4)
- Method: DARE-TIES
- Genome:
- Architecture: (35.95B total, 3B active)
- Size: 71.9GB (bfloat16)
Darwin v4
Produced by the Darwin v4 evolutionary LLM merging platform (A2AP).
Platform: Phase 1 proxy fitness search + Phase 2 DARE-TIES tensor merge.
Inference Test
Verified: code generation ✓, Korean ✓, English ✓
Load time: ~19s on 2×H200, ~11 tokens/sec
