--- license: apache-2.0 language: - en - ko - zh - ja - multilingual library_name: transformers pipeline_tag: text-generation tags: - darwin - darwin-v9 - darwin-jgos - moe - mixture-of-experts - reasoning - gpqa - benchmark - greedy - vidraft - eval-results base_model: - Qwen/Qwen3.5-397B-A17B base_model_relation: merge model-index: - name: Darwin-397B-JGOS results: - task: type: text-generation name: Graduate-Level Reasoning dataset: name: GPQA Diamond type: Idavidrein/gpqa config: gpqa_diamond split: train metrics: - type: accuracy value: 90.9 name: Accuracy (greedy, single-sample, no test-time engine) verified: false --- # Darwin-397B-JGOS โ€” Darwin V9 Platform ยท 397B MoE ยท GPQA 90.9 % (Pure Greedy)

GPQA REASON

Opus 36B

Family FINAL Bench

> Largest Darwin model ยท Qwen 3.5 397B base + Darwin V9 FFN transplant ยท 397B MoE (~17B active) ยท BF16 > **GPQA Diamond: 90.9 % โ€” pure greedy, single-sample, NO test-time engine** --- ## Overview **Darwin-397B-JGOS** is the largest and highest-scoring member of the Darwin family. Built on **Qwen 3.5 397B** as the base, it transplants the FFN (expert) strengths of multiple high-performance models through the **Darwin V9 platform**, producing a 397B-parameter Mixture-of-Experts model with ~17B active parameters per token. It reaches **90.9 % on GPQA Diamond with pure greedy decoding (single sample)** โ€” surpassing **Darwin-28B-REASON (89.39 %, achieved *with* the Darwin-DELPHI test-time engine)** without using any test-time engine at all. This is the highest GPQA Diamond score in the Darwin family to date. --- ## ๐Ÿงฌ Darwin Platform & Research **Darwin** is VIDRAFT's measuring-result-driven reasoning model family โ€” approximately **20 official models** plus **400+ community derivatives**, ranking among the top open models on GPQA. - **Darwin V9 platform** โ€” evolutionary FFN/expert transplant and trust-weighted merging onto large-scale MoE backbones. - **FINAL Bench** โ€” VIDRAFT's evaluation framework. - **4-layer Pre-AGI roadmap** โ€” Darwin โ†’ AETHER โ†’ PROMETHEUS โ†’ HEPHAESTUS. --- ## ๐Ÿงฌ Model Lineage | Role | Model | Contribution | |:---:|:---|:---| | **Base** | `Qwen 3.5 397B (A17B)` | 397B Mixture-of-Experts backbone (~17B active). | | **FFN transplant** | **Darwin V9 platform** (proprietary) | Transplants the FFN (expert) strengths of multiple high-performance models onto the base. | | **Result** | **`Darwin-397B-JGOS`** (this model) | 397B MoE โ†’ **90.9 %** GPQA Diamond, pure greedy. | > The full Darwin V9 merge recipe โ€” source models, weighting, and density โ€” is **proprietary** and **not disclosed** (trade secret). --- ## โš™๏ธ Technical Specifications | Component | Value | |:---|:---| | Architecture | `Qwen3_5MoeForConditionalGeneration` (Qwen 3.5 generation MoE) | | Parameters | **~397 B total / ~17 B active** (Mixture-of-Experts) | | Base | Qwen 3.5 397B (A17B) | | Precision | bfloat16 | | License | other | --- ## ๐Ÿ”ฌ Core Technique โ€” Darwin V9 Platform Darwin V9 transplants the FFN (expert) strengths of multiple high-performance models onto a Qwen 3.5 397B MoE base, then applies trust-weighted evolutionary merging. > The source models, merge weights, and density schedule are **proprietary** and constitute a **trade secret**; they 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 | Standard | 88.89 % (176 / 198) | | Darwin-28B-REASON | Darwin-DELPHI (test-time) | 89.39 % (177 / 198) | | **Darwin-397B-JGOS** | **Greedy (single-sample, no engine)** | **๐Ÿฅ‡ 90.9 % (180 / 198)** | **Reproducible evaluation settings:** - Greedy decoding (temperature = 0), single sample โ€” **no voting / self-consistency / test-time engine** - Max generation: 16,384 tokens - Answer options shuffled (seed = 42) - Hardware: **NVIDIA B200** (tensor-parallel 2 ร— pipeline-parallel 3, 6 GPUs) - Inference engine: **vLLM**, bfloat16, `max_model_len = 18432` > Darwin-397B-JGOS achieves the family's top GPQA Diamond score using nothing but greedy decoding โ€” no Darwin-DELPHI, no majority voting. --- ## ๐Ÿš€ Usage (vLLM) ```bash vllm serve FINAL-Bench/Darwin-397B-JGOS --tensor-parallel-size 2 --pipeline-parallel-size 3 --dtype bfloat16 --trust-remote-code ``` --- ## ๐ŸŽฏ Recommended Use-Cases - Graduate-level STEM reasoning (GPQA / science qualifying exams) - Mathematical problem solving - Complex multi-step chain-of-thought - Code generation and debugging - Bilingual reasoning (strong English + Korean; also Chinese / Japanese) ## โš ๏ธ Limitations - 397B MoE in bfloat16 requires multi-GPU serving (e.g. B200 ร—6 with TP2ร—PP3). - The 90.9 % figure is a single-run greedy measurement on GPQA Diamond (198 items). - Reasoning traces can be verbose โ€” control with max tokens. --- ## ๐Ÿ“š Citation ```bibtex @misc{darwin397b_jgos_2026, title = {Darwin-397B-JGOS: Darwin V9 Platform FFN Transplant on a 397B MoE Base}, author = {FINAL-Bench / Darwin Research Team}, year = {2026}, howpublished = {https://huggingface.co/FINAL-Bench/Darwin-397B-JGOS}, note = {Darwin V9 - 90.9 percent GPQA Diamond (greedy, single-sample)} } ``` --- ## ๐Ÿ”— Related Darwin Models - **Darwin-28B-REASON** โ€” RTD + Darwin-DELPHI, GPQA 89.39 % - **Darwin-28B-Opus** โ€” base, GPQA 88.89 % (HF-official GPQA top tier) - **Darwin-36B-Opus** โ€” MoE 36B, GPQA 88.4 % - **Darwin-27B-Opus** โ€” 27B dense, GPQA 86.9 % - **Darwin-9B-NEG** โ€” 9B Negentropy, GPQA 84.3 % --- *Darwin-397B-JGOS ยท Darwin V9 Platform ยท 90.9 % GPQA Diamond (pure greedy) ยท FINAL-Bench*