FINAL-Bench / Darwin-28B-REASON

Darwin-28B-REASON — Reasoning-Trace Distilled, Darwin-DELPHI Enhanced

Full standalone reasoning model derived from Darwin-28B-Opus · Reasoning-Trace Distillation (RTD) · Darwin-DELPHI test-time engine · 27.6 B · BF16 · Apache 2.0 GPQA Diamond: 89.39 % with Darwin-DELPHI

---

Overview

Darwin-28B-REASON is a reasoning-enhanced standalone model derived from Darwin-28B-Opus . It combines two components:

1. Reasoning-Trace Distillation (RTD) — a reasoning-trace distillation stage applied on top of the Darwin-28B-Opus base, producing this fully self-contained model (full weights, no external adapter required).
2. Darwin-DELPHI — a proprietary test-time reasoning engine.

Together they push graduate-level scientific reasoning to the top tier of the Darwin family: 89.39 % on GPQA Diamond with Darwin-DELPHI. The model is released under Apache-2.0 .

---

🧬 Darwin Platform & Research

Darwin is VIDRAFT's measuring-result-driven Korean reasoning model family — approximately 20 official models plus 400+ community derivatives , ranking #3 globally on GPQA among open models. The base model, Darwin-28B-Opus , is the HuggingFace-official GPQA #3 (88.89 %) model.

• Platform technique — MRI trust-weighted Evolutionary Merge ( arXiv:2605.14386 ).
• FINAL Bench — VIDRAFT's evaluation framework (SSRN): MetaCognition +14.05 , MA-ER Gap 0.392 .
• 4-layer Pre-AGI roadmap — Darwin → AETHER → PROMETHEUS → HEPHAESTUS.

---

🧬 Model Lineage

---

⚙️ Technical Specifications

---

🔬 Core Techniques

① RTD — Reasoning-Trace Distillation

RTD distills complete reasoning chains from a publicly available mathematical corpus (Apache-2.0 source) on top of the Darwin-28B-Opus base, producing this standalone model. It strengthens long-form, multi-step scientific reasoning while preserving the base model's bilingual capability.

The full RTD recipe (curation, trace selection, training schedule) is proprietary and is not disclosed.

② Darwin-DELPHI — Test-Time Reasoning Engine

Darwin-DELPHI is a proprietary test-time engine applied at inference. It performs multi-sample cross-validation , re-examination of uncertain responses , and iterative self-critique , converging to a consensus answer through a single-agent Delphi-method procedure.

Darwin-DELPHI is not stored in the model weights . Its internal parameters — sampling counts, stage transitions, and decision thresholds — are a trade secret and are not published.

---

🏆 Benchmark — GPQA Diamond (198 questions)

GPQA Diamond is a 198-question, PhD-level graduate science reasoning benchmark.

The evaluation methodology for the Darwin-DELPHI result is protected ; sample counts, staging, and thresholds are a trade secret .

---

🚀 Usage

Darwin-28B-REASON is a full standalone model — load it directly, no base model or adapter merge required.

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

MODEL = "FINAL-Bench/Darwin-28B-REASON"

tok = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model.eval()

messages = [
    {"role": "user",
     "content": "A particle moves along x(t) = t³ − 6t² + 9t. Find when it is at rest and classify the motion."}
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048)
print(tok.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))

The 89.39 % GPQA Diamond result is produced with the Darwin-DELPHI test-time engine applied on top of this model. Darwin-DELPHI is provided through the Darwin-series evaluation harness.

---

🎯 Recommended Use-Cases

• Graduate-level STEM reasoning (GPQA / science qualifying exams)
• Mathematical problem solving (MATH, AIME-style problems)
• Complex multi-step chain-of-thought tasks
• Code generation and debugging
• Bilingual reasoning (strong English + Korean; also Chinese / Japanese)

⚠️ Limitations

• The 27.6 B model in bfloat16 requires ≈ 55 GB of VRAM (a single A100-80GB or B200 is sufficient).
• The 89.39 % result depends on the Darwin-DELPHI test-time engine; the model on its own delivers strong but lower single-model accuracy.
• Optimised for English first, with secondary support for Korean, Chinese, and Japanese.
• Reasoning traces tend to be verbose — control with max_new_tokens as needed.

---

📚 Citation

@misc{darwin28b_reason_2026,
  title  = {Darwin-28B-REASON: Reasoning-Trace Distillation and Darwin-DELPHI Test-Time Reasoning on Darwin-28B-Opus},
  author = {FINAL-Bench / Darwin Research Team},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/FINAL-Bench/Darwin-28B-REASON}},
  note   = {RTD + Darwin-DELPHI · 89.39 % GPQA Diamond}
}

@misc{darwin_family_2026,
  title  = {Darwin Family: MRI Trust-Weighted Evolutionary Merging for Reasoning Models},
  author = {VIDRAFT / FINAL-Bench},
  year   = {2026},
  howpublished = {\url{https://arxiv.org/abs/2605.14386}}
}

@misc{final_bench_2026,
  title  = {FINAL Bench: A Measuring-Result-Driven Evaluation Framework for Reasoning Models},
  author = {VIDRAFT / FINAL-Bench},
  year   = {2026},
  howpublished = {SSRN}
}

---

🔗 Related Darwin Models

• Darwin-28B-Opus — base model, Qwen3.6-27B × Opus distilled, GPQA 88.89 %
• Darwin-36B-Opus — MoE 36B, GPQA 88.4 %
• Darwin-27B-Opus — 27B dense (Qwen3.5 generation), GPQA 86.9 %
• Darwin-9B-NEG — 9B with Negentropy distillation, GPQA 84.3 %
• Darwin-4B-Genesis — smallest Darwin member

---

This model is introduced in Darwin Family .

Darwin-28B-REASON · RTD + Darwin-DELPHI · 89.39 % GPQA Diamond · FINAL-Bench