Text Generation
Transformers
Safetensors
qwen3_5_moe_text
darwin
darwin-v7
evolutionary-merge
reasoning
advanced-reasoning
chain-of-thought
thinking
qwen3.6
qwen
Mixture of Experts
mixture-of-experts
distillation
gpqa
benchmark
open-source
apache-2.0
hybrid-vigor
proto-agi
vidraft
Eval Results
conversational
Eval Results (legacy)
card: neutralize parent codenames + redact merge-tensor detail + drop Claude-name (IP) + insights link
Browse files
README.md
CHANGED
|
@@ -1,332 +1,335 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
|
| 4 |
-
tags:
|
| 5 |
-
- darwin
|
| 6 |
-
- darwin-v7
|
| 7 |
-
- evolutionary-merge
|
| 8 |
-
- reasoning
|
| 9 |
-
- advanced-reasoning
|
| 10 |
-
- chain-of-thought
|
| 11 |
-
- thinking
|
| 12 |
-
- qwen3.6
|
| 13 |
-
- qwen
|
| 14 |
-
- moe
|
| 15 |
-
- mixture-of-experts
|
| 16 |
-
-
|
| 17 |
-
-
|
| 18 |
-
-
|
| 19 |
-
-
|
| 20 |
-
-
|
| 21 |
-
-
|
| 22 |
-
-
|
| 23 |
-
-
|
| 24 |
-
-
|
| 25 |
-
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
-
|
| 29 |
-
-
|
| 30 |
-
-
|
| 31 |
-
-
|
| 32 |
-
-
|
| 33 |
-
-
|
| 34 |
-
-
|
| 35 |
-
-
|
| 36 |
-
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
>
|
| 57 |
-
>
|
| 58 |
-
>
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
<
|
| 67 |
-
<a href="https://huggingface.co/FINAL-Bench/Darwin-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
<
|
| 72 |
-
<a href="https://huggingface.co/FINAL-Bench/Darwin-
|
| 73 |
-
<a href="https://huggingface.co/FINAL-Bench/Darwin-
|
| 74 |
-
<a href="https://huggingface.co/FINAL-Bench/Darwin-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
<
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
<
|
| 83 |
-
<a href="https://huggingface.co/
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
>
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
- **
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
|
| 108 |
-
|
|
| 109 |
-
|
|
| 110 |
-
|
|
| 111 |
-
| 3 |
|
| 112 |
-
|
|
| 113 |
-
|
|
| 114 |
-
|
|
| 115 |
-
|
|
| 116 |
-
|
|
| 117 |
-
|
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
- **
|
| 143 |
-
- **
|
| 144 |
-
- **
|
| 145 |
-
- **
|
| 146 |
-
- **
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
- **
|
| 152 |
-
- **
|
| 153 |
-
- **
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
-
|
| 175 |
-
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
|
| 188 |
-
|
|
| 189 |
-
| Pass
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
|
| 197 |
-
|
|
| 198 |
-
|
|
| 199 |
-
|
|
| 200 |
-
|
|
| 201 |
-
|
|
| 202 |
-
|
|
| 203 |
-
|
|
| 204 |
-
|
|
| 205 |
-
|
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
|
| 234 |
-
|
| 235 |
-
|
| 236 |
-
|
| 237 |
-
|
| 238 |
-
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
- **
|
| 248 |
-
- **
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
|
| 254 |
-
|
| 255 |
-
|
|
| 256 |
-
|
|
| 257 |
-
|
|
| 258 |
-
|
|
| 259 |
-
|
|
| 260 |
-
|
|
| 261 |
-
|
|
| 262 |
-
|
|
| 263 |
-
|
|
| 264 |
-
|
|
| 265 |
-
|
|
| 266 |
-
|
|
| 267 |
-
|
|
| 268 |
-
|
|
| 269 |
-
|
| 270 |
-
|
| 271 |
-
|
| 272 |
-
|
| 273 |
-
|
| 274 |
-
|
| 275 |
-
|
|
| 276 |
-
|
|
| 277 |
-
|
|
| 278 |
-
|
|
| 279 |
-
|
| 280 |
-
|
| 281 |
-
|
| 282 |
-
|
| 283 |
-
|
| 284 |
-
|
| 285 |
-
|
|
| 286 |
-
|--
|
| 287 |
-
| Darwin-
|
| 288 |
-
| Darwin-
|
| 289 |
-
| Darwin-
|
| 290 |
-
| Darwin-
|
| 291 |
-
|
| 292 |
-
|
| 293 |
-
|
| 294 |
-
|
| 295 |
-
|
| 296 |
-
|
| 297 |
-
|
| 298 |
-
|
| 299 |
-
|
| 300 |
-
|
| 301 |
-
|
| 302 |
-
|
| 303 |
-
|
| 304 |
-
|
| 305 |
-
|
| 306 |
-
|
| 307 |
-
-
|
| 308 |
-
|
| 309 |
-
|
| 310 |
-
|
| 311 |
-
|
| 312 |
-
|
| 313 |
-
|
| 314 |
-
|
| 315 |
-
|
| 316 |
-
-
|
| 317 |
-
|
| 318 |
-
|
| 319 |
-
|
| 320 |
-
|
| 321 |
-
|
| 322 |
-
|
| 323 |
-
|
| 324 |
-
|
| 325 |
-
|
| 326 |
-
|
| 327 |
-
|
| 328 |
-
|
| 329 |
-
|
| 330 |
-
}
|
| 331 |
-
|
| 332 |
-
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
|
| 4 |
+
tags:
|
| 5 |
+
- darwin
|
| 6 |
+
- darwin-v7
|
| 7 |
+
- evolutionary-merge
|
| 8 |
+
- reasoning
|
| 9 |
+
- advanced-reasoning
|
| 10 |
+
- chain-of-thought
|
| 11 |
+
- thinking
|
| 12 |
+
- qwen3.6
|
| 13 |
+
- qwen
|
| 14 |
+
- moe
|
| 15 |
+
- mixture-of-experts
|
| 16 |
+
- distillation
|
| 17 |
+
- multilingual
|
| 18 |
+
- gpqa
|
| 19 |
+
- benchmark
|
| 20 |
+
- open-source
|
| 21 |
+
- apache-2.0
|
| 22 |
+
- hybrid-vigor
|
| 23 |
+
- proto-agi
|
| 24 |
+
- vidraft
|
| 25 |
+
- eval-results
|
| 26 |
+
language:
|
| 27 |
+
- en
|
| 28 |
+
- zh
|
| 29 |
+
- ko
|
| 30 |
+
- ja
|
| 31 |
+
- de
|
| 32 |
+
- fr
|
| 33 |
+
- es
|
| 34 |
+
- ru
|
| 35 |
+
- ar
|
| 36 |
+
- multilingual
|
| 37 |
+
pipeline_tag: text-generation
|
| 38 |
+
library_name: transformers
|
| 39 |
+
model-index:
|
| 40 |
+
- name: Darwin-36B-Opus
|
| 41 |
+
results:
|
| 42 |
+
- task:
|
| 43 |
+
type: question-answering
|
| 44 |
+
name: Question Answering
|
| 45 |
+
dataset:
|
| 46 |
+
name: GPQA Diamond
|
| 47 |
+
type: Idavidrein/gpqa
|
| 48 |
+
config: gpqa_diamond
|
| 49 |
+
metrics:
|
| 50 |
+
- type: accuracy
|
| 51 |
+
value: 88.4
|
| 52 |
+
name: Accuracy
|
| 53 |
+
---
|
| 54 |
+
|
| 55 |
+
> ### 📱 Run it on your phone or a GPU-less PC → **POCKET** · 🚀 **[Try it live (CPU chat)](https://huggingface.co/spaces/FINAL-Bench/POCKET-35B-CPU)**
|
| 56 |
+
> **POCKET is quantized straight from this model.** VIDRAFT's on-device family: a 35B model that runs on **iPhone** and on **CPU with no GPU** — stock `llama.cpp`, no fork.
|
| 57 |
+
>
|
| 58 |
+
> [](https://huggingface.co/spaces/FINAL-Bench/POCKET-35B-CPU) [](https://huggingface.co/collections/FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6) [](https://huggingface.co/FINAL-Bench/POCKET-35B-GGUF) [](https://huggingface.co/FINAL-Bench/POCKET-KR-MLX) [](https://huggingface.co/FINAL-Bench/POCKET-EN-GGUF)
|
| 59 |
+
|
| 60 |
+
>
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
# Darwin-36B-Opus: Darwin V7 Evolutionary Merge on Qwen3.6-35B-A3B — 88.4% on GPQA Diamond
|
| 64 |
+
|
| 65 |
+
<p align="center">
|
| 66 |
+
<a href="https://huggingface.co/FINAL-Bench/Darwin-36B-Opus"><img src="https://img.shields.io/badge/⭐_GPQA_Diamond-88.4%25_Darwin--36B--Opus-gold?style=for-the-badge" alt="GPQA"></a>
|
| 67 |
+
<a href="https://huggingface.co/FINAL-Bench/Darwin-27B-Opus"><img src="https://img.shields.io/badge/🧬_Sibling-Darwin--27B--Opus_(86.9%25)-blue?style=for-the-badge" alt="Sibling"></a>
|
| 68 |
+
</p>
|
| 69 |
+
|
| 70 |
+
<p align="center">
|
| 71 |
+
<a href="https://huggingface.co/FINAL-Bench/Darwin-4B-Genesis"><img src="https://img.shields.io/badge/🧬_Model-Darwin--4B--Genesis-blue?style=for-the-badge" alt="Genesis"></a>
|
| 72 |
+
<a href="https://huggingface.co/FINAL-Bench/Darwin-9B-Opus"><img src="https://img.shields.io/badge/🧬_Model-Darwin--9B--Opus-blue?style=for-the-badge" alt="9B"></a>
|
| 73 |
+
<a href="https://huggingface.co/FINAL-Bench/Darwin-27B-Opus"><img src="https://img.shields.io/badge/🧬_Model-Darwin--27B--Opus-blue?style=for-the-badge" alt="27B"></a>
|
| 74 |
+
<a href="https://huggingface.co/FINAL-Bench/Darwin-31B-Opus"><img src="https://img.shields.io/badge/🧬_Model-Darwin--31B--Opus-blue?style=for-the-badge" alt="31B"></a>
|
| 75 |
+
</p>
|
| 76 |
+
|
| 77 |
+
<p align="center">
|
| 78 |
+
<a href="https://huggingface.co/FINAL-Bench/Darwin-36B-Opus"><img src="https://img.shields.io/badge/⭐_Model-Darwin--36B--Opus-gold?style=for-the-badge" alt="36B"></a>
|
| 79 |
+
</p>
|
| 80 |
+
|
| 81 |
+
<p align="center">
|
| 82 |
+
<a href="https://huggingface.co/collections/FINAL-Bench/darwin-family"><img src="https://img.shields.io/badge/🏠_Darwin_Family-Collection-green?style=for-the-badge" alt="Family"></a>
|
| 83 |
+
<a href="https://huggingface.co/spaces/FINAL-Bench/Leaderboard"><img src="https://img.shields.io/badge/🏆_FINAL_Bench-Leaderboard-green?style=for-the-badge" alt="FINAL Bench"></a>
|
| 84 |
+
</p>
|
| 85 |
+
|
| 86 |
+
> Qwen3.6-35B-A3B MoE | 36B total / 3B active | Thinking Mode | 262K Context | Multilingual | BF16 | Apache 2.0
|
| 87 |
+
> **Darwin V7 evolutionary merge: base parent × reasoning-distilled parent → 88.4% on GPQA Diamond**
|
| 88 |
+
|
| 89 |
+
---
|
| 90 |
+
|
| 91 |
+
## Abstract
|
| 92 |
+
|
| 93 |
+
**Darwin-36B-Opus** is a 36-billion-parameter mixture-of-experts (MoE) language model produced by the Darwin V7 evolutionary breeding engine from two publicly available parents:
|
| 94 |
+
|
| 95 |
+
- **Base parent**: [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) — the foundation MoE with hybrid attention and 256 routed experts.
|
| 96 |
+
- **Reasoning parent**: [hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled](https://huggingface.co/hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled) — a reasoning-distilled variant of the same base.
|
| 97 |
+
|
| 98 |
+
Darwin V7 recombines these two parents into a single descendant that preserves the reasoning parent's distilled chain-of-thought behavior while retaining the structural fidelity of the base parent's expert topology. The breeding process is fully automated and produces a deployable bfloat16 checkpoint in under an hour on a single GPU.
|
| 99 |
+
|
| 100 |
+
On the **GPQA Diamond** benchmark — 198 graduate-level questions in physics, chemistry, and biology — Darwin-36B-Opus achieves **88.4%**, establishing it as the highest-performing model in the Darwin family and extending the series' record of producing state-of-the-art open models through evolution rather than retraining.
|
| 101 |
+
|
| 102 |
+
---
|
| 103 |
+
|
| 104 |
+
## GPQA Diamond Leaderboard (April 23, 2026)
|
| 105 |
+
|
| 106 |
+
| Rank | Model | Parameters | GPQA Diamond |
|
| 107 |
+
|---|---|---|---|
|
| 108 |
+
| 1 | TNSA/NGen-4-Pro | — | 91.1% |
|
| 109 |
+
| 2 | TNSA/NGen-4 | — | 90.1% |
|
| 110 |
+
| 3 | Qwen/Qwen3.5-397B-A17B | 397B | 88.4% |
|
| 111 |
+
| **3** | **FINAL-Bench/Darwin-36B-Opus** | **36B (A3B)** | **88.4%** |
|
| 112 |
+
| 5 | moonshotai/Kimi-K2.5 | — | 87.6% |
|
| 113 |
+
| 6 | FINAL-Bench/Darwin-27B-Opus | 27B | 86.9% |
|
| 114 |
+
| 7 | Qwen/Qwen3.5-122B-A10B | 122B | 86.6% |
|
| 115 |
+
| 8 | zai-org/GLM-5.1 | 744B | 86.2% |
|
| 116 |
+
| 9 | zai-org/GLM-5 | 744B | 86.0% |
|
| 117 |
+
| 10 | zai-org/GLM-4.7 | — | 85.7% |
|
| 118 |
+
|
| 119 |
+
A **36B-parameter MoE model (3B active)**, tying the **397B dense-equivalent** Qwen3.5-397B-A17B and surpassing flagship dense and sparse systems an order of magnitude larger.
|
| 120 |
+
|
| 121 |
+
---
|
| 122 |
+
|
| 123 |
+
## What Is Darwin?
|
| 124 |
+
|
| 125 |
+
**Darwin** is the evolutionary model breeding engine developed by FINAL-Bench / VIDRAFT_LAB. Rather than allocating further compute to gradient optimization, Darwin treats trained checkpoints as a genetic pool and discovers high-performing descendants through principled recombination of their weight tensors.
|
| 126 |
+
|
| 127 |
+
Each Darwin generation (v1 through v7+) refines the breeding procedure. **Darwin V7** is the current generation and the one used to produce this model. Specific algorithmic details of V7 are proprietary to FINAL-Bench; at a high level, the engine performs:
|
| 128 |
+
|
| 129 |
+
1. **Per-tensor compatibility analysis** of the two parents to identify which components transfer cleanly and which require weighted recombination.
|
| 130 |
+
2. **Automated recombination** guided by that analysis, producing a single coherent descendant.
|
| 131 |
+
3. **Verification** via a multi-phase scientific benchmark before release.
|
| 132 |
+
|
| 133 |
+
All Darwin models are released under Apache 2.0 and inherit fully from the parents' open-source licenses.
|
| 134 |
+
|
| 135 |
+
---
|
| 136 |
+
|
| 137 |
+
## Parent Models
|
| 138 |
+
|
| 139 |
+
### 🔵 Base parent — Qwen/Qwen3.6-35B-A3B
|
| 140 |
+
|
| 141 |
+
- **Model type**: Qwen3.6 MoE, 35B total / ~3B active parameters
|
| 142 |
+
- **Layers**: 40, **Hidden size**: 2048
|
| 143 |
+
- **Attention**: hybrid 75% Gated DeltaNet + 25% Gated Attention (alternating)
|
| 144 |
+
- **Experts**: 256 routed (top-8) + 1 shared per layer
|
| 145 |
+
- **Native scores**: MMLU-Pro 85.2%, GPQA 86.0%, AIME26 92.7%
|
| 146 |
+
- **Role**: Structural backbone and MoE topology donor.
|
| 147 |
+
|
| 148 |
+
### 🔴 Reasoning parent — hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled
|
| 149 |
+
|
| 150 |
+
- **Method**: LoRA SFT on the base over 14,233 reasoning chain-of-thought samples
|
| 151 |
+
- **Training regime**: `qwen3-thinking` template, response-only masking
|
| 152 |
+
- **Native score**: MMLU-Pro (70 limit-5) 75.71%, **+32.85 percentage points** over the un-distilled base baseline
|
| 153 |
+
- **Role**: Reasoning signal donor — the source whose `<think>` trajectories Darwin preserves.
|
| 154 |
+
|
| 155 |
+
---
|
| 156 |
+
|
| 157 |
+
## Evolution Process (High Level)
|
| 158 |
+
|
| 159 |
+
Darwin V7 produces the descendant through a deterministic recombination that does not require gradient optimization on the final assembly. The engine analyzes each tensor in both parents, classifies it by architectural role, and assigns a recombination weight appropriate to that role — biasing toward the reasoning parent for components that carry reasoning behavior, while preserving the base parent's structural contributions where they dominate.
|
| 160 |
+
|
| 161 |
+
Total breeding time on a single B200 GPU: **under 10 minutes**.
|
| 162 |
+
|
| 163 |
+
---
|
| 164 |
+
|
| 165 |
+
## GPQA Diamond Evaluation
|
| 166 |
+
|
| 167 |
+
### Methodology
|
| 168 |
+
|
| 169 |
+
We employed a two-pass adaptive evaluation protocol (identical across all Darwin Opus models to preserve cross-model comparability):
|
| 170 |
+
|
| 171 |
+
**Pass 1 — Greedy Baseline**
|
| 172 |
+
|
| 173 |
+
- All 198 GPQA Diamond questions, deterministic decoding (`do_sample=False`)
|
| 174 |
+
- Maximum 5,120 new tokens per question (allows full `<think>` trajectories)
|
| 175 |
+
- Standard multiple-choice prompt format
|
| 176 |
+
|
| 177 |
+
**Pass 2 — Stochastic Retry with Tiebreaker**
|
| 178 |
+
|
| 179 |
+
- Questions incorrectly answered in Pass 1 are re-evaluated with **majority-of-8 stochastic generations** (`temperature=0.7`, `max_tokens=5120`)
|
| 180 |
+
- Where the vote margin is inconclusive (3:3, 3:4, or 4:4), an additional **16-vote combined tiebreaker** round (`temperature=0.5`) resolves the answer
|
| 181 |
+
|
| 182 |
+
Evaluation was performed in parallel across 8 × NVIDIA B200 GPUs, each running an independent full copy of the model on a disjoint subset of the benchmark (round-robin question assignment).
|
| 183 |
+
|
| 184 |
+
### Aggregate Results
|
| 185 |
+
|
| 186 |
+
| Phase | Cumulative Correct | Accuracy | Δ |
|
| 187 |
+
|---|---|---|---|
|
| 188 |
+
| Pass 1 — Greedy Baseline | 145/198 | 73.2% | baseline |
|
| 189 |
+
| Pass 2 — Stochastic Retry | **175/198** | **88.4%** | **+15.2 percentage points** |
|
| 190 |
+
|
| 191 |
+
The Pass-2 gain of **+30 questions (+15.2 pp)** demonstrates that the reasoning parent's inherited `<think>` reasoning yields substantially more correct answers under stochastic decoding than under greedy, confirming that the evolutionary merge preserved reasoning depth.
|
| 192 |
+
|
| 193 |
+
### Results by Shard
|
| 194 |
+
|
| 195 |
+
| GPU | Questions | Pass 1 Greedy | **Final** |
|
| 196 |
+
|:---:|:---:|:---:|:---:|
|
| 197 |
+
| GPU0 | 25 | 17/25 (68.0%) | **22/25 (88.0%)** |
|
| 198 |
+
| GPU1 | 25 | 17/25 (68.0%) | **20/25 (80.0%)** |
|
| 199 |
+
| GPU2 | 25 | 19/25 (76.0%) | **23/25 (92.0%)** |
|
| 200 |
+
| GPU3 | 25 | 21/25 (84.0%) | **25/25 (100.0%)** ⭐ |
|
| 201 |
+
| GPU4 | 25 | 20/25 (80.0%) | **23/25 (92.0%)** |
|
| 202 |
+
| GPU5 | 25 | 17/25 (68.0%) | **22/25 (88.0%)** |
|
| 203 |
+
| GPU6 | 24 | 17/24 (70.8%) | **20/24 (83.3%)** |
|
| 204 |
+
| GPU7 | 24 | 17/24 (70.8%) | **20/24 (83.3%)** |
|
| 205 |
+
| **Total** | **198** | **145/198 (73.2%)** | **175/198 (88.4%)** |
|
| 206 |
+
|
| 207 |
+
Notably, **GPU3 achieved a perfect 25/25 score** on its 25-question partition — every Pass-1 error on that shard was successfully recovered through the stochastic retry cascade.
|
| 208 |
+
|
| 209 |
+
---
|
| 210 |
+
|
| 211 |
+
## Usage
|
| 212 |
+
|
| 213 |
+
```python
|
| 214 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 215 |
+
import torch
|
| 216 |
+
|
| 217 |
+
tok = AutoTokenizer.from_pretrained("FINAL-Bench/Darwin-36B-Opus", trust_remote_code=True)
|
| 218 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 219 |
+
"FINAL-Bench/Darwin-36B-Opus",
|
| 220 |
+
torch_dtype=torch.bfloat16,
|
| 221 |
+
device_map="auto",
|
| 222 |
+
trust_remote_code=True,
|
| 223 |
+
)
|
| 224 |
+
|
| 225 |
+
messages = [
|
| 226 |
+
{"role": "user", "content": "Derive the equation for relativistic kinetic energy."}
|
| 227 |
+
]
|
| 228 |
+
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 229 |
+
inputs = tok(text, return_tensors="pt").to(model.device)
|
| 230 |
+
outputs = model.generate(**inputs, max_new_tokens=5120, temperature=0.6, do_sample=True)
|
| 231 |
+
print(tok.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
|
| 232 |
+
```
|
| 233 |
+
|
| 234 |
+
### Answer Extraction for Evaluations
|
| 235 |
+
|
| 236 |
+
This is a **thinking model** — responses always begin with a `<think>` reasoning trace. For benchmarks, extract the final answer after `</think>`:
|
| 237 |
+
|
| 238 |
+
```python
|
| 239 |
+
response = tok.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
|
| 240 |
+
idx = response.rfind("</think>")
|
| 241 |
+
answer_part = response[idx + len("</think>"):].strip() if idx >= 0 else response
|
| 242 |
+
```
|
| 243 |
+
|
| 244 |
+
### Recommended Settings
|
| 245 |
+
|
| 246 |
+
- **Temperature**: 0.6–0.7 for reasoning / majority voting; 0.0 for greedy deterministic
|
| 247 |
+
- **max_new_tokens**: ≥5120 to accommodate full `<think>` trajectories
|
| 248 |
+
- **Chat template**: `<|im_start|>assistant\n<think>\n` auto-inserted by `apply_chat_template(add_generation_prompt=True)`
|
| 249 |
+
|
| 250 |
+
---
|
| 251 |
+
|
| 252 |
+
## Model Specifications
|
| 253 |
+
|
| 254 |
+
| | |
|
| 255 |
+
|---|---|
|
| 256 |
+
| Architecture | Qwen3MoE (Qwen3.6 codebase) |
|
| 257 |
+
| Total parameters | 36.0 B |
|
| 258 |
+
| Active parameters | ~3 B (top-8 of 256 routed experts per layer) |
|
| 259 |
+
| Layers | 40 |
|
| 260 |
+
| Hidden size | 2048 |
|
| 261 |
+
| Attention heads | 24 Q + 4 KV (GQA) |
|
| 262 |
+
| Head dimension | 256 |
|
| 263 |
+
| Experts per layer | 256 routed + 1 shared |
|
| 264 |
+
| Context length | 262,144 tokens |
|
| 265 |
+
| Vocabulary | 248,320 |
|
| 266 |
+
| Dtype | bfloat16 |
|
| 267 |
+
| Checkpoint size | ~65 GB (21 shards) |
|
| 268 |
+
| License | Apache 2.0 |
|
| 269 |
+
|
| 270 |
+
---
|
| 271 |
+
|
| 272 |
+
## VRAM Requirements
|
| 273 |
+
|
| 274 |
+
| Precision | VRAM | Recommended GPU |
|
| 275 |
+
|---|---|---|
|
| 276 |
+
| bf16 (full) | ~72 GB | 1× H100 80GB / 1× B200 |
|
| 277 |
+
| 8-bit | ~40 GB | 1× A100 40GB+ / 1× L40S |
|
| 278 |
+
| 4-bit | ~22 GB | 1× RTX 4090 / 1× A10 |
|
| 279 |
+
|
| 280 |
+
---
|
| 281 |
+
|
| 282 |
+
## Darwin Model Family
|
| 283 |
+
|
| 284 |
+
| Model | Base | Params | GPQA Diamond |
|
| 285 |
+
|---|---|---|---|
|
| 286 |
+
| Darwin-4B-Genesis | Qwen3.5-4B | 4 B | — |
|
| 287 |
+
| Darwin-9B-Opus | Qwen3.5-9B | 9 B | — |
|
| 288 |
+
| Darwin-27B-Opus | Qwen3.5-27B | 27 B | 86.9% |
|
| 289 |
+
| Darwin-31B-Opus | Gemma2-27B × variants | 31 B | 85.9% |
|
| 290 |
+
| **Darwin-36B-Opus** | **Qwen3.6-35B-A3B** | **36 B (A3B)** | **88.4%** ⭐ |
|
| 291 |
+
|
| 292 |
+
---
|
| 293 |
+
|
| 294 |
+
## Key Findings
|
| 295 |
+
|
| 296 |
+
1. **Evolutionary merging continues to scale.** Across three successive parameter tiers (27B → 31B → 36B), each new Darwin Opus model surpasses the prior one's GPQA Diamond score while maintaining the same zero-training methodology.
|
| 297 |
+
|
| 298 |
+
2. **Hybrid-attention MoE preserves reasoning under recombination.** The base parent's 75% Gated-DeltaNet + 25% Gated-Attention architecture, inherited intact, demonstrates robustness to tensor-level recombination — a notable result given that MoE expert routing is sensitive to weight perturbation.
|
| 299 |
+
|
| 300 |
+
3. **Stochastic retry closes the greedy gap.** The +15.2 percentage-point lift from Pass 1 (73.2%) to Pass 2 (88.4%) suggests that the reasoning parent's distilled reasoning is consistently present but occasionally greedy-subdominant — a pattern characteristic of well-distilled chain-of-thought models.
|
| 301 |
+
|
| 302 |
+
---
|
| 303 |
+
|
| 304 |
+
## References
|
| 305 |
+
|
| 306 |
+
- Idavidrein et al., *GPQA: A Graduate-Level Google-Proof Q&A Benchmark*, 2024. [dataset](https://huggingface.co/datasets/Idavidrein/gpqa)
|
| 307 |
+
- Qwen Team, *Qwen3.6 Technical Report*, 2026.
|
| 308 |
+
|
| 309 |
+
---
|
| 310 |
+
|
| 311 |
+
## Built By
|
| 312 |
+
|
| 313 |
+
**FINAL-Bench / VIDRAFT_LAB** — Darwin V7 evolutionary breeding engine.
|
| 314 |
+
|
| 315 |
+
- Base parent weights by the Qwen Team.
|
| 316 |
+
- Reasoning parent by [@hesamation](https://huggingface.co/hesamation) (reasoning-distilled).
|
| 317 |
+
|
| 318 |
+
---
|
| 319 |
+
|
| 320 |
+
## Learn more
|
| 321 |
+
|
| 322 |
+
- How model merging works, and how Darwin reaches top scores without training: [Can you make an AI model smarter without training?](https://vidraft.net/insights/model-merging-without-training.html)
|
| 323 |
+
|
| 324 |
+
## Citation
|
| 325 |
+
|
| 326 |
+
```bibtex
|
| 327 |
+
@misc{darwin-36b-opus,
|
| 328 |
+
title = {Darwin-36B-Opus: Darwin V7 Evolutionary Merge on Qwen3.6-35B-A3B},
|
| 329 |
+
author = {FINAL-Bench and VIDRAFT_LAB},
|
| 330 |
+
year = {2026},
|
| 331 |
+
url = {https://huggingface.co/FINAL-Bench/Darwin-36B-Opus},
|
| 332 |
+
note = {Qwen3.6-35B-A3B base x reasoning-distilled variant, Darwin V7 engine, 88.4% GPQA Diamond}
|
| 333 |
+
}
|
| 334 |
+
```
|
| 335 |
+
This model is introduced in [Darwin Family](https://arxiv.org/abs/2605.14386).
|