Post
657
๐งฌ Darwin-35B-A3B-Opus โ The Child That Surpassed Both Parents
What if a merged model could beat both its parents? We proved it can.
Darwin-35B-A3B-Opus is a 35B MoE model (3B active) built with our Darwin V5 engine โ the first evolution system that CT-scans parent models before merging them.
๐ค Model: FINAL-Bench/Darwin-35B-A3B-Opus
The result speaks for itself: GPQA Diamond 90.0%, versus Father (Qwen3.5-35B-A3B) at 84.2% and Mother (Claude 4.6 Opus Distilled) at 85.0%. That's +6.9% over Father and +5.9% over Mother. Not a tradeoff โ a genuine leap. Meanwhile, MMMLU sits at 85.0% (Father: 85.2%), multimodal is fully intact, and all 201 languages are preserved.
How? Model MRI changed everything. Traditional merging is guesswork. Darwin V4 added evolution. Darwin V5 added X-ray vision. Model MRI scans each parent layer by layer and discovers: Mother's L34โL38 is the reasoning engine (peak cosine distance), 50โ65% of Mother's experts are dead (killed by text-only distillation), and Father is a healthy generalist with every expert alive. The prescription: transplant Mother's reasoning brain at L38 (90% weight), replace her dead experts with Father's living ones, and let Father's router handle the output layer. Reasoning went up. Versatility stayed intact. No tradeoff โ just evolution.
35B total, 3B active (MoE) ยท GPQA Diamond 90.0% ยท MMMLU 85.0% (201 languages) ยท Multimodal Image & Video ยท 262K native context ยท 147.8 tok/s on H100 ยท Runs on a single RTX 4090 (Q4) ยท Apache 2.0
Darwin V5's full algorithm and technical details will be released alongside an upcoming paper.
๐ Live Demo: FINAL-Bench/Darwin-35B-A3B-Opus
๐ FINAL Bench Leaderboard: FINAL-Bench/Leaderboard
๐ ALL Bench Leaderboard: FINAL-Bench/all-bench-leaderboard
Built by VIDRAFT ยท Supported by the Korean Government GPU Support Program
What if a merged model could beat both its parents? We proved it can.
Darwin-35B-A3B-Opus is a 35B MoE model (3B active) built with our Darwin V5 engine โ the first evolution system that CT-scans parent models before merging them.
๐ค Model: FINAL-Bench/Darwin-35B-A3B-Opus
The result speaks for itself: GPQA Diamond 90.0%, versus Father (Qwen3.5-35B-A3B) at 84.2% and Mother (Claude 4.6 Opus Distilled) at 85.0%. That's +6.9% over Father and +5.9% over Mother. Not a tradeoff โ a genuine leap. Meanwhile, MMMLU sits at 85.0% (Father: 85.2%), multimodal is fully intact, and all 201 languages are preserved.
How? Model MRI changed everything. Traditional merging is guesswork. Darwin V4 added evolution. Darwin V5 added X-ray vision. Model MRI scans each parent layer by layer and discovers: Mother's L34โL38 is the reasoning engine (peak cosine distance), 50โ65% of Mother's experts are dead (killed by text-only distillation), and Father is a healthy generalist with every expert alive. The prescription: transplant Mother's reasoning brain at L38 (90% weight), replace her dead experts with Father's living ones, and let Father's router handle the output layer. Reasoning went up. Versatility stayed intact. No tradeoff โ just evolution.
35B total, 3B active (MoE) ยท GPQA Diamond 90.0% ยท MMMLU 85.0% (201 languages) ยท Multimodal Image & Video ยท 262K native context ยท 147.8 tok/s on H100 ยท Runs on a single RTX 4090 (Q4) ยท Apache 2.0
Darwin V5's full algorithm and technical details will be released alongside an upcoming paper.
๐ Live Demo: FINAL-Bench/Darwin-35B-A3B-Opus
๐ FINAL Bench Leaderboard: FINAL-Bench/Leaderboard
๐ ALL Bench Leaderboard: FINAL-Bench/all-bench-leaderboard
Built by VIDRAFT ยท Supported by the Korean Government GPU Support Program
