They Merged Two AI Models — The Child Came Out Smarter Than Both Parents

Darwin-35B-A3B-Opus
Darwin-35B-A3B-Opus

Model Space FINAL Bench ALL Bench

"The child surpassed both parents — that is evolution."

TL;DR: 35B MoE (3B active) | GPQA Diamond 90.0% (vs Father 84.2% & Mother 85.0%) | MMMLU 85.0% | Multimodal ✅ | 201 Languages | 262K Context | 147.8 tok/s | Apache 2.0

Table of Contents
Why Darwin — The Child That Surpassed Both Parents
Model Overview
Parent Models
Darwin V5 — Beyond Simple Merging
Model MRI Scans — Parent Neural Anatomy
Child Model Health Check — MRI Verification
Inherited Capabilities
Father's Official Benchmarks (Reference)
Performance & Hardware Requirements
Model Specifications
Usage
Built By
FAQ

1. Why Darwin — The Child That Surpassed Both Parents There is a fundamental question at the heart of AI model merging: If the parent models already exist, why crossbreed at all?

This model is the answer.

Benchmark Results
GPQA Diamond (198 Questions, Graduate-Level Reasoning)

Model Accuracy Multimodal Benchmark Published
🧬 Darwin-35B-A3B-Opus (Child) 90.0% ✅ Image/Video ✅ Fully Open
👩 Mother — Jackrong Claude 4.6 Opus Distilled 85.0% ❌ Text-only ❌ Not Published
👨 Father — Qwen3.5-35B-A3B (Official) 84.2% ✅ Image/Video ✅ Official
Evaluation: SGLang, context 32768, temperature 0, greedy decoding, official GPQA prompt format ("ANSWER: LETTER")

MMMLU (Multilingual Knowledge, 29 Languages)

Model Accuracy
🧬 Darwin-35B-A3B-Opus (Child) 85.0%
👨 Father — Qwen3.5-35B-A3B (Official) 85.2%
Darwin preserves Father-level multilingual knowledge while achieving decisively superior reasoning.

The child outperformed both parents in reasoning and matched the Father in multilingual knowledge.

GPQA vs Father: +6.9% relative improvement ((90.0−84.2)/84.2)
GPQA vs Mother: +5.9% relative improvement ((90.0−85.0)/85.0)
MMMLU: 85.0% — Father-level (85.2%) multilingual knowledge preserved
Why Not Simply Use the Mother?
Mother (Claude Distilled) Darwin (Child)
Reasoning Strong (85.0%) Stronger (90.0%)
Image/Video ❌ Lost during text-only fine-tuning ✅ Inherited from Father
201 Languages ❌ Potentially degraded ✅ Inherited from Father
262K Context Unverified ✅ Father's architecture preserved
Benchmark Transparency ❌ No scores published ✅ Fully open
Why Not Simply Use the Father?
The Father (Qwen3.5-35B-A3B) excels in versatility but plateaus at 84.2% on hard reasoning tasks. Darwin pushes reasoning to 90.0% while retaining Father-level multilingual knowledge (MMMLU 85.0% vs 85.2%) along with all general-purpose capabilities.

Bottom line: Darwin is the only model that exceeds the Mother's reasoning, preserves the Father's multilingual knowledge, and retains full multimodal capability — all at once.

1. Model Overview Darwin-35B-A3B-Opus is a next-generation reasoning-enhanced language model produced by VIDRAFT's Darwin V5 evolution engine.

Darwin V5 fuses two key innovations:

Evolutionary Merge — Applies natural selection to automatically discover optimal weight combinations across generations of candidates
Model MRI Integration — CT-scans each parent model layer by layer before merging, steering the evolutionary process with structural insight
If conventional merging is "mixing ingredients blindfolded," Darwin V5 is "precision surgery under X-ray guidance."

1. Parent Models Role Model Strengths 👨 Father Qwen/Qwen3.5-35B-A3B General knowledge, multimodal (image/video), coding, agents, 201 languages, 262K context 👩 Mother Jackrong/Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled Claude 4.6 Opus CoT distillation, structured step-by-step reasoning, coding agent compatibility
2. Darwin V5 — Beyond Simple Merging The Limitations of Conventional Merging Traditional model merging requires humans to set hyperparameters — ratio, density, and the like — by intuition. You pick ratio=0.5, density=0.9, run the merge once, and hope for the best. The outcome hinges on luck, and applying a single ratio uniformly across billions of parameters ignores the distinct role each layer plays.

Darwin V4's Breakthrough
Darwin V4 addressed this with evolutionary algorithms — automatically exploring hundreds of parameter combinations and selecting survivors based on real benchmark scores. Yet V4 was still blind evolution: it had no understanding of what each layer actually does.

Darwin V5: Model MRI Opens the Eyes
V5 integrates Model MRI — a neural anatomy analyzer — to give the evolutionary process "sight":

[Phase 0] Model MRI — CT-scan both parents, layer by layer
↓ "Father's layers 15–25 concentrate multilingual knowledge"
↓ "Mother's layers 30–40 concentrate reasoning patterns"
↓
[Phase 1] MRI-Guided Evolution — Begin from a scan-informed initial genome
↓ Not random, but "initialized from CT findings"
↓
[Phase 2] mergekit real merge + benchmark-driven fitness selection
↓ Faster convergence within the MRI-narrowed search space
↓
[Phase 3] MRI Health Check — CT-scan the child model
↓ Detect interference and function loss
↓ Prescribe layer-specific ratio adjustments
↓
[Final] Darwin-35B-A3B-Opus

V4 vs V5 at a Glance
Darwin V4 Darwin V5
Analogy Mixing ingredients blindfolded Precision surgery under X-ray
Initial genome Random MRI-guided
Layer control 2 ratios (attn/ffn) 40 layers independently
Pre-diagnosis ❌ None ✅ Phase 0 MRI scan
Post-verification Benchmark only ✅ Phase 3 health check
Search efficiency Broad, unfocused Narrowed, guided search
Failure diagnosis Unknown "why" Pinpoints the failing layer
Darwin V4: Discovered Parameters (Blind Evolution)
Parameter Value Interpretation
ratio 0.481 Father 52% : Mother 48% — asymmetric blend
density_a 0.855 85.5% of Father's weights selected
density_b 0.971 97.1% of Mother's weights adopted
attn 0.168 Only 16.8% modification in attention layers
ffn 0.841 84.1% modification in FFN layers
What this means: Attention patterns (determining what to focus on) are almost entirely preserved from the Father, while FFN layers (the knowledge store) are largely overwritten with the Mother's reasoning patterns.

Discovering attn=0.168 alongside ffn=0.841 — this extreme asymmetry — is virtually impossible to arrive at through human intuition.

Darwin V5: The MRI-Guided Merge Recipe
After scanning both parents, Model MRI prescribed a fundamentally different recipe:

MRI-Guided Genome

Parameter V4 (Blind) V5 (MRI) Shift
global_ratio 0.481 0.800 Mother weight ↑↑
attn_ratio 0.168 0.320 Attention also shifts toward Mother
ffn_ratio 0.841 0.590 FFN becomes more conservative
density_a 0.855 0.799 Similar
density_b 0.971 0.799 Mother density ↓ (Dead Expert compensation)
The key insight: MRI prescribed "draw more heavily from the Mother (ratio 0.8), but reduce density (0.799) because 50–65% of her experts are dead." V4, searching blindly, landed on ratio=0.481 — the opposite direction entirely.

Layer-Wise Merge Strategy (3 Surgical Blocks)
MRI did not prescribe uniform ratios. Instead, it partitioned all 40 layers into 3 distinct blocks:

Merge Ratio + Parent Importance + MoE Health per Layer

Block Layers t (Mother %) Router Source Rationale
Block 1 L0–L37 59.9% Mother Reasoning pattern injection across the bulk of the network
Block 2 L38 90.0% Mother Golden Layer — the Mother's core reasoning engine
Block 3 L39 53.4% Father Output layer — Father's router preserves multimodal routing
L38 is the "Golden Layer": The Mother's MRI revealed peak cosine distance at L34–L38 (see Mother MRI below). Darwin V5 responded by assigning t=0.9 to L38 — transplanting the Mother's reasoning engine nearly in its entirety.

1. Model MRI Scans — Parent Neural Anatomy Mother MRI: Claude 4.6 Opus Distilled Mother Probe Cosine Distance

Probe-wise Layer Importance: Layers L34–L38 light up in intense red (high cosine distance) across the REASONING, CODE, and LOGIC probes — this is the Mother's reasoning engine.

Mother MoE Health

Metric Status Interpretation
Router Entropy ✅ ~1.0 across all layers Healthy — experts are evenly distributed
Dead Expert % 🔴 50–65% Critical — Claude distillation killed half the experts
Expert Similarity ✅ 0.001–0.008 Healthy — surviving experts remain diverse
A Dead Expert rate of 50–65% is the telltale fingerprint of Claude's text-only distillation. The fine-tuning process silenced multimodal and multilingual experts that were never activated during text-only training.

Mother Expert Utilization Heatmap

Expert Utilization Heatmap: The map is predominantly dark (inactive), with only sparse bright activations — the Claude reasoning pattern is concentrated in a small cluster of specialized experts.

Father MRI: A Healthy Generalist (The Organ Donor)
Father MoE Health

Father Expert Utilization Heatmap

Father Layer Importance by Probe

The Father (Qwen3.5-35B-A3B) exhibits healthy, uniform expert activation across all 40 layers — a well-balanced generalist with every expert alive and contributing. He serves as the "organ donor" who revives the Mother's dead 50–65% of experts.

Parent Comparison: Layer Advantage Map
Parent A vs B Layer Advantage

Above zero (↑ A): Father is stronger — primarily L0–L5 (embedding and early layers)
Below zero (↓ B): Mother is stronger — scattered but consistent from L5 through L35
L34–L38: Mother shows her strongest advantage on the REASONING and CODE probes
L39: Father recovers — the output layer favors Father's multimodal routing
This advantage map directly informed the 3-block merge recipe: Mother dominates L0–L38, Father reclaims L39.

How GPQA 90% Was Achieved
Mother L34–L38: reasoning engine (MRI red zone)
↓ t=0.9 — transplanted nearly in full
+
Father L39: output router (multimodal/multilingual expert activation)
↓ t=0.53 — Father's routing preserved
+
Dead Expert replacement → Father's living experts fill the Mother's dead slots
↓
= GPQA 90.0% (surpassing both parents)

The Mother's "reasoning brain" was transplanted while her dead experts were replaced with the Father's living counterparts. Reasoning went up; versatility stayed intact.

Evolution History
Phase 1 → Phase 2 evolution complete
Final real_score: 0.8405
Merge time: 181.6 seconds
Merge commit: 109838c2

1. Child Model Health Check — MRI Verification Darwin Health Check — Child vs Parents

✅ Verdict: Healthy — No issues detected.

The chart above plots the layer-by-layer importance of the child (Darwin, green bars) against both parents (Father = blue dashed, Mother = red dashed). Key findings:

Layer 0 (Embedding): The child's importance spikes to 0.42 — both parents exhibit similar peaks (~0.35–0.50). The child has successfully inherited the critical embedding layer from both parents with no interference.

Layers 1–33 (Middle): Near-zero importance across all three models. This is expected — middle layers in MoE architectures process information incrementally, with no single layer acting as a bottleneck. The child tracks both parents precisely, confirming zero function loss across the bulk of the network.

Layers 34–39 (Reasoning Engine): Importance rises sharply. This is the exact region where the Mother's MRI revealed intense reasoning activity (cosine distance > 0.6). The child's green bars match or exceed both parents — demonstrating that the Mother's reasoning patterns were successfully transplanted while the Father's output routing was preserved.

Layer 39 (Output): The child peaks at ~0.48, closely tracking both parents. The final output layer is intact.

Why This Matters
The MRI health check confirms three critical outcomes:

No interference — There is no layer where the child's importance abnormally exceeds the parents' (which would signal weight conflict)
No function loss — There is no layer where the parents had high importance but the child collapsed to zero
Successful transplant — The L34–L39 reasoning engine from the Mother is fully operational in the child
Darwin V5 MRI-Guided Merge Recipe

MRI-guided layer-wise merge (3 blocks)

Genome: ratio=0.800 attn=0.320 ffn=0.590 density=0.799

L0–L37: t=0.5988 (Mother 60%) — router from Mother
L38: t=0.9000 (Mother 90%) — "Golden Layer" reasoning core
L39: t=0.5336 (Father 47%) — router from Father (output routing)

Insight Detail
L38 = "Golden Layer" MRI identified L34–L38 as the Mother's reasoning core. Darwin assigned t=0.9 (90% Mother) to L38 specifically
Router Strategy: B→B→A Mother's router for the reasoning layers, Father's router for the final output — preserving both the reasoning pathways and multimodal routing
Dead Expert Revival The Mother's 50–65% dead experts (killed during text-only fine-tuning) were replaced with the Father's living experts — restoring multimodal and multilingual capabilities
📄 The full algorithm and technical details of the Darwin V5 evolution engine will be released alongside an upcoming paper.

1. Inherited Capabilities From the Father (Qwen3.5-35B-A3B) Multimodal: Image and video understanding 201 Languages: Global linguistic coverage 262K Context: Native long-context support (extendable to 1M via YaRN) Gated DeltaNet + MoE: Efficient hybrid architecture Multi-Token Prediction: Improved inference throughput From the Mother (Claude 4.6 Opus Distilled) Structured Thinking: Systematic step-by-step reasoning within tags Efficient Reasoning: "Let me analyze this request carefully: 1… 2… 3…" pattern Coding Agent Compatibility: Native "developer" role support for Claude Code and OpenCode Tool Calling Stability: Consistent performance in tool-use scenarios Autonomous Execution: Extended autonomous operation in agentic environments
2. Father's Official Benchmarks (Reference) Darwin is built on this architecture with enhanced reasoning:

Category Benchmark Father Official
Knowledge MMLU-Pro 85.3
Knowledge MMLU-Redux 93.3
Reasoning GPQA Diamond 84.2
Reasoning HLE w/ CoT 22.4
Math HMMT Feb 2025 89.0
Coding SWE-bench Verified 69.2
Coding LiveCodeBench v6 74.6
Agent TAU2-Bench 81.2
Agent BFCL-V4 (Tool Use) 67.3
Instruction IFEval 91.9
Multilingual MMMLU 85.2
Agentic Search BrowseComp 61.0

1. Performance & Hardware Requirements
   Inference Speed
   Metric Value
   Generation Speed 147.8 tok/s
   Environment Single NVIDIA H100 93GB NVL, SGLang, BF16
   Qwen Official API 162.8 tok/s (Alibaba Cloud)
   Hardware Requirements
   Setup VRAM Status
   BF16 (Full Precision) 65.5 GiB
   
   Single H100 93GB NVL 93 GB ✅ Comfortable
   Single A100 80GB 80 GB ⚠️ Tight
   Single A100 40GB 40 GB ❌ Insufficient
   Q8 Quantized ~35 GiB
   Single A100 40GB 40 GB ✅ Feasible
   Q4_K_M Quantized ~18 GiB
   Single RTX 4090 24GB 24 GB ✅ Comfortable
   2× RTX 4090 (tp=2) 48 GB ✅ BF16 feasible
   As a Mixture-of-Experts model, only 3B parameters are active per token despite loading the full 35B. This sparsity means quantization has minimal impact on output quality.

2. Model Specifications
   Architecture Qwen3.5 MoE (Gated DeltaNet + MoE)
   Total Parameters 35B
   Active Parameters 3B per forward pass
   Hidden Dimension 2,048
   Layers 40
   Layer Layout 10 × (3 × GDN→MoE + 1 × Attention→MoE)
   Experts 256 (8 routed + 1 shared active)
   Expert Intermediate Dim 512
   Context Length 262,144 native (up to 1,010,000 via YaRN)
   Languages 201
   Multimodal ✅ Image & Video input
   License Apache 2.0
   Engine Darwin V5 (Evolutionary Merge + Model MRI)
   Evolution Phase Phase 2, real_score 0.8405
   Merge Commit 109838c2

3. Usage
   SGLang (Recommended)
   python -m sglang.launch_server \
   --model-path FINAL-Bench/Darwin-35B-A3B-Opus \
   --tp 1 \
   --mem-fraction-static 0.90 \
   --context-length 32768 \
   --trust-remote-code

vLLM
vllm serve FINAL-Bench/Darwin-35B-A3B-Opus \
--trust-remote-code \
--enforce-eager

Transformers
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("FINAL-Bench/Darwin-35B-A3B-Opus", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
"FINAL-Bench/Darwin-35B-A3B-Opus",
dtype="bfloat16",
device_map="auto",
trust_remote_code=True,
)

Best Practices
Use context ≥ 32K for reasoning tasks — the model leverages extended thinking
For maximum reasoning quality, use thinking mode (default) with generous max_tokens (≥ 16384)
The model generates blocks for internal reasoning; extract the final answer after

1. Built By Developer VIDRAFT Evolution Engine Darwin V5 (Evolutionary Merge + Model MRI) Infrastructure 4 × NVIDIA H100 93GB NVL GPU Merge Time 181.6 seconds Shard Distribution 14 shards → GPU [1, 2, 3] round-robin Acknowledgements Korean Government — This research was supported by the Korean Government's 'GPU Support Program' research grant Qwen Team — Qwen3.5-35B-A3B base architecture Jackrong — Claude 4.6 Opus Reasoning Distilled model nohurry, TeichAI — Distillation datasets Citation @misc{vidraft_darwin_35b_opus, title = {Darwin-35B-A3B-Opus: MRI-Guided Evolutionary Merge Beyond Both Parents}, author = {VIDRAFT}, year = {2026}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/FINAL-Bench/Darwin-35B-A3B-Opus}} }

Contact
📧 [e-mail]

1. FAQ What is Darwin-35B-A3B-Opus? How does Darwin V5 differ from simple model merging? What GPU do I need to run this model? Does it support multimodal inputs (images/video)? What languages does it support? What is Model MRI? What are "Dead Experts" and why do they matter? Is this model open source? #DarwinAI #EvolutionaryMerge #ModelMRI #DarwinV5 #GPQA90 #Qwen35 #MoE3B #Reasoning #Multimodal #201Languages #OpenSource #Apache2 #VIDRAFT #NaturalSelection #LayerWiseMerge #ClaudeOpus #ThinkingModel #CodingAgent #LongContext262K #BestOpenSourceLLM2026 #DeadExpertRevival #GoldenLayer #MoEMerge #NeuralAnatomy