--- license: apache-2.0 language: - en - zh - ko - ja - multilingual library_name: transformers pipeline_tag: text-generation tags: - darwin - darwin-reason - reasoning - advanced-reasoning - chain-of-thought - thinking - reasoning-trace-distillation - rtd - darwin-delphi - test-time-compute - qwen3.6 - qwen - gpqa - benchmark - open-source - apache-2.0 - proto-agi - vidraft - eval-results base_model: - FINAL-Bench/Darwin-28B-Opus base_model_relation: finetune model-index: - name: Darwin-28B-REASON results: - task: type: text-generation name: Graduate-Level Reasoning dataset: type: Idavidrein/gpqa name: GPQA Diamond config: gpqa_diamond split: train metrics: - type: accuracy value: 89.39 name: Accuracy (with Darwin-DELPHI) verified: false --- # Darwin-28B-REASON โ€” Reasoning-Trace Distilled, Darwin-DELPHI Enhanced

GPQA Opus

36B 27B NEG

Family FINAL Bench

> Full standalone reasoning model derived from Darwin-28B-Opus ยท Reasoning-Trace Distillation (RTD) ยท Darwin-DELPHI test-time engine ยท 27.6 B ยท BF16 ยท 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](https://huggingface.co/FINAL-Bench/Darwin-28B-Opus)**. It combines two components: 1. **Reasoning-Trace Distillation (RTD)** โ€” 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** โ€” 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**. --- ## ๐Ÿงฌ Darwin Platform & Research **Darwin** is VIDRAFT's measuring-result-driven Korean reasoning model family โ€” 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** โ€” MRI trust-weighted Evolutionary Merge ([arXiv:2605.14386](https://arxiv.org/abs/2605.14386)). - **FINAL Bench** โ€” VIDRAFT's evaluation framework (SSRN): MetaCognition **+14.05**, MA-ER Gap **0.392**. - **4-layer Pre-AGI roadmap** โ€” Darwin โ†’ AETHER โ†’ PROMETHEUS โ†’ HEPHAESTUS. --- ## ๐Ÿงฌ Model Lineage | Role | Model | Contribution | |:---:|:---|:---| | **Base** | [`FINAL-Bench/Darwin-28B-Opus`](https://huggingface.co/FINAL-Bench/Darwin-28B-Opus) | GPQA #3 (88.89 %) Qwen3.6-generation reasoning backbone. | | **RTD training** | reasoning-trace distillation | Distills complete reasoning chains into the model on top of the Opus base. | | **Test-time engine** | Darwin-DELPHI | Proprietary inference-time consensus engine (not stored in weights). | | **Result** | **`Darwin-28B-REASON`** (this model) | Full standalone RTD model + Darwin-DELPHI โ†’ **89.39 %** GPQA Diamond. | --- ## โš™๏ธ Technical Specifications | Component | Value | |:---|:---| | Architecture | `Qwen3_5ForConditionalGeneration` (Qwen3.6 generation, hybrid linear + full attention; text path, `language_model_only`) | | Parameters | **27.6 B** (BF16) โ€” full standalone weights | | Layers | 64 (3 linear : 1 full attention, `full_attention_interval = 4`) | | Vocab size | 248 320 | | Context length | 262 144 (long-chain reasoning supported) | | Delivery | Full self-contained model โ€” no external base or adapter required | | Precision | bfloat16 | | License | Apache 2.0 | --- ## ๐Ÿ”ฌ Core Techniques ### โ‘  RTD โ€” 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. ### โ‘ก Darwin-DELPHI โ€” 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 โ€” sampling counts, stage transitions, and decision thresholds โ€” are a **trade secret** and are not published. --- ## ๐Ÿ† Benchmark โ€” GPQA Diamond (198 questions) GPQA Diamond is a 198-question, PhD-level graduate science reasoning benchmark. | Model | Engine | **Accuracy** | |:---|:---|:---:| | Darwin-28B-Opus (base) | Standard | 88.89 % (176 / 198) | | **Darwin-28B-REASON** | **Darwin-DELPHI** | **๐Ÿฅ‡ 89.39 % (177 / 198)** | The evaluation methodology for the Darwin-DELPHI result is **protected**; sample counts, staging, and thresholds are a **trade secret**. --- ## ๐Ÿš€ Usage Darwin-28B-REASON is a **full standalone model** โ€” load it directly, no base model or adapter merge required. ```python 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) = tยณ โˆ’ 6tยฒ + 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. --- ## ๐ŸŽฏ 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) ## โš ๏ธ Limitations - The 27.6 B model in bfloat16 requires โ‰ˆ 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 โ€” control with `max_new_tokens` as needed. --- ## ๐Ÿ“š Citation ```bibtex @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 ยท 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} } ``` --- ## ๐Ÿ”— Related Darwin Models - **Darwin-28B-Opus** โ€” base model, Qwen3.6-27B ร— Opus distilled, GPQA 88.89 % - **Darwin-36B-Opus** โ€” MoE 36B, GPQA 88.4 % - **Darwin-27B-Opus** โ€” 27B dense (Qwen3.5 generation), GPQA 86.9 % - **Darwin-9B-NEG** โ€” 9B with Negentropy distillation, GPQA 84.3 % - **Darwin-4B-Genesis** โ€” smallest Darwin member --- This model is introduced in [Darwin Family](https://arxiv.org/abs/2605.14386). *Darwin-28B-REASON ยท RTD + Darwin-DELPHI ยท 89.39 % GPQA Diamond ยท FINAL-Bench*