Darwin-31B-Opus

Gemma 4 Dense 31B | Thinking Mode | 256K Context | 140+ Languages | BF16 | Apache 2.0

---

Overview

Darwin-31B-Opus is a reasoning-enhanced model created by merging google/gemma-4-31B-it (Father) and TeichAI/gemma-4-31B-it-Claude-Opus-Distill (Mother) using the Darwin V6 engine.

Darwin V6 diagnoses both parent models at the tensor level before merging, assigning an independent optimal ratio to each of the 1,188 tensors. This is fundamentally different from conventional merging tools that apply a single uniform ratio across all tensors.

---

Parent Models

Role	Model	Characteristics
Father	google/gemma-4-31B-it	Gemma 4 Dense 31B, multimodal, 256K context, LMArena 1452 (open model #3)
Mother	TeichAI/gemma-4-31B-it-Claude-Opus-Distill	Claude 4.6 Opus high-effort reasoning distillation, code/science/analysis

Model Diagnostic Scan (MDS)

Left: Father (gemma-4-31B-it) — balanced generalist with low activation across most probes. Right: Mother (Claude-Opus-Distill) — strong REASONING concentration in L50-L60, CODE activation in late layers, KOREAN at start and end. The Mother shows significantly more specialized layer patterns from Claude Opus distillation.

---

Benchmarks

Benchmark	Darwin-31B-Opus	Father (gemma-4-31B-it)	Condition
ARC-Challenge	82.89%	-	loglikelihood, zero-shot, 200Q
GPQA Diamond	66.0%	60.0%	generative thinking mode, greedy, 50Q

GPQA Diamond was evaluated under identical conditions for both models: same 50 questions, same seed (i+42), same prompt template, greedy decoding (do_sample=False), max_new_tokens=2048, enable_thinking=True. Darwin-31B-Opus achieved a 10% relative improvement over the Father model.

Note: Gemma 4 architecture (Gemma4ForConditionalGeneration) has limited compatibility with lm-eval's loglikelihood method due to its multimodal wrapper structure. Only generative evaluation produces valid results for Gemma 4 based models. Full 198-question evaluation with Majority Voting is planned.

---

Darwin V6 vs Conventional Merging

Capability	mergekit (DARE-TIES)	Darwin V6
Implementation	Library call (mergekit CLI)	Direct PyTorch tensor operations, no external dependency
Ratio selection	Uniform ratio across all tensors	Per-tensor ratio from MDS diagnostic (1,188 independent ratios)
Pre-merge analysis	None	Static tensor profiling (entropy, std, norm) + probe-based functional importance (5 probes)
Ratio formula	Human-set or grid search	combined = static × 0.4 + probe × 0.6, then evolutionary optimization
Transplant	Not supported	ratio < 0.15 → Father 100%, ratio > 0.85 → Mother 100% (zero interpolation noise)
Post-merge validation	Benchmark score only	Layer-by-layer Health Check: child vs both parents, interference and function loss detection
Search method	Manual tuning	CMA-ES evolution with adaptive 14-dimensional genome
Reproducibility	Config file	genome_hash seed guarantees identical output for identical genome
GPU efficiency	Single merge per run	Phase 1 proxy (200 steps, seconds) → Phase 2 real merge (top-k only evaluated)

---

How Darwin V6 Works

Darwin V6 does not use mergekit or any external merge library. It re-implements DARE-TIES (Yadav et al., 2023) directly via PyTorch tensor operations with per-tensor diagnostic ratios.

Before merging, Darwin performs a Model Diagnostic Scan (MDS) on both parents. For every tensor, it measures Shannon entropy (information density), standard deviation (activation spread), and L2 norm (energy). Additionally, 5 diagnostic probes (REASONING, CODE, MATH, KNOWLEDGE, LANGUAGE) are passed through the model, measuring cosine distance when each layer is skipped to determine functional importance.

The final merge ratio for each tensor:

static_score = entropy × 0.3 + std × 0.2 + clamp(norm, 100) × 0.002
probe_score  = Σ(cosine_distance[probe_i] × weight_i)
combined     = static × 0.4 + probe × 0.6
mri_ratio    = combined_b / (combined_a + combined_b)
final_ratio  = mri_ratio × mri_trust + genome_ratio × (1 - mri_trust)

The mri_trust parameter itself is optimized by the CMA-ES evolutionary algorithm, allowing the system to automatically determine the optimal balance between diagnostic prescription and evolutionary search for each model pair.

After merging, a Health Check compares the child model against both parents layer-by-layer, detecting interference (child importance >> parent max) or function loss (parent importance high but child dropped).

Parent Comparison (MDS Result)

---

Evolution Result

Best Score (ARC-Challenge)	0.8289
Merge Method	DARE-TIES (direct PyTorch)
Tensors Merged	1,188
Health Check	healthy
Phase 2 Steps	4 (early stop, patience=5)
Total Time	134 min
Infrastructure	4 x NVIDIA H100 NVL (100GB)

Optimal Genome (14-dimensional adaptive):

global_ratio:        0.5147   (overall merge ratio)
attn_ratio:          0.3169   (Attention layers — Father dominant)
ffn_ratio:           0.9316   (FFN layers — Mother dominant)
embed_ratio:         0.7748   (Embedding)
density_a:           0.8997   (Father DARE density)
density_b:           0.9539   (Mother DARE density)
block_0_ratio:       0.6628   (L0-L9)
block_1_ratio:       0.6431   (L10-L19)
block_2_ratio:       0.5146   (L20-L29, balanced)
block_3_ratio:       0.5971   (L30-L39)
block_4_ratio:       0.6339   (L40-L49)
block_5_ratio:       0.8583   (L50-L59, reasoning core — Mother dominant)
mri_trust:           0.3631   (MDS 36% + Genome 64%)
merge_method_weight: 0.6897

Key observations from the genome: ffn_ratio=0.93 indicates the FFN layers strongly favor the Mother (Claude Opus Distill), and block_5 (L50-L59)=0.86 shows the reasoning core layers also favor Mother. This aligns with the MDS heatmap pattern where Mother's reasoning capability concentrated in the final layers. Meanwhile, attn_ratio=0.32 preserves Father's attention structure, maintaining the original Gemma 4 multimodal and long-context capabilities.

---

Model Specifications

Architecture	Gemma 4 Dense (Hybrid Attention: Sliding Window + Global)
Parameters	31B
Precision	BF16
Context	256,072
Languages	140+
Thinking	enable_thinking=True chain-of-thought
License	Apache 2.0

---

Usage

Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("FINAL-Bench/Darwin-31B-Opus", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "FINAL-Bench/Darwin-31B-Opus",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "Prove that sqrt(2) is irrational."}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=True
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

---

VRAM Requirements

Setup	VRAM	Status
BF16 Full Precision	~62 GB
NVIDIA H100 80GB	80 GB	Single GPU
NVIDIA A100 80GB x 2	160 GB	Comfortable
NVIDIA RTX 4090 24GB x 4	96 GB	device_map=auto

---

References

• DARE-TIES: Yadav et al., 2023 (https://arxiv.org/abs/2311.03099) — re-implemented, not library-dependent
• Darwin V6 Engine: https://huggingface.co/spaces/ginigen-ai/DARWIN-V5-BACKUP
• FINAL Bench: https://huggingface.co/spaces/FINAL-Bench/Leaderboard

---

Built By

Developer	VIDRAFT
Engine	Darwin V6 (Diagnostic-Guided Evolutionary Merge)
Architecture	Gemma-4-31B
License	Apache 2.0

---

Citation

@misc{vidraft_darwin_31b_opus,
  title        = {Darwin-31B-Opus: Diagnostic-Guided Evolutionary Merge on Gemma 4},
  author       = {VIDRAFT},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/FINAL-Bench/Darwin-31B-Opus}}
}