Darwin-180B-RSI-R3: VIDRAFT's Self-Improving MoE Model Hits #2 in Math & Science Reasoning Globally

Darwin-180B-RSI-R3: VIDRAFT's Self-Improving MoE Model Hits #2 in Math & Science Reasoning Globally TL;DR: VIDRAFT's Darwin-180B-RSI-R3 is a 180B-parameter open-weight MoE model that applies Recursive Self-Improvement (RSI) to autonomously boost its own reasoning weights — no new human-labeled data required. It scored #2 in math & science reasoning on SuperCLUE's November 2026 evaluation, outperforming models from Google, DeepSeek, and xAI. The weights are publicly available on Hugging Face with over 4.68 million cumulative downloads. What it is Darwin-180B-RSI-R3 is VIDRAFT's latest open-weight large language model built on a Mixture-of-Experts (MoE) architecture. Key characteristics: Total parameters: 180 billion Active parameters at inference: approximately 3 billion (sparse MoE activation) Base model: Alibaba's Qwen3.8-Flash-Next Training approach: Recursive Self-Improvement (RSI) applied on top of the base model's existing knowledge-storage expert layers License: Open-weight, commercially usable Compressed variant: POCKET-Darwin-180B — a quantized/distilled version designed to run on consumer CPUs (personal PCs and laptops, no GPU required) This is the only Korean model included in the November 2026 SuperCLUE evaluation across 19 major US and Chinese AI models. How it works The core technical contribution is Recursive Self-Improvement (RSI) — a training paradigm that eliminates the need for continuously feeding the model hand-labeled answer data to improve performance. At a high level, the RSI loop works like this: Problem generation: The model autonomously generates new reasoning problems within a target domain (mathematics, science). Self-solution: The model attempts to solve the problems it generated. Self-verification: The model evaluates and critiques its own solutions. Weight update: Only the reasoning-related weights are updated based on verified solutions. The existing expert layers responsible for general knowledge storage remain frozen. This design is architecturally deliberate. By preserving the knowledge expert layers from the Qwen3.8 base and directing self-improvement only at reasoning pathways, VIDRAFT avoids catastrophic forgetting of broad world knowledge while still compounding gains on targeted reasoning tasks. The process iterates over multiple rounds (the "R3" suffix denotes the third RSI round), allowing performance to compound without human annotation at each step. The sparse MoE structure also matters for efficiency: activating only ~3B of 180B parameters per forward pass keeps inference costs manageable despite the model's large total capacity. Benchmarks & results All numbers below are sourced directly from the SuperCLUE November 2026 evaluation and VIDRAFT's public announcement. SuperCLUE Math & Science Reasoning (out of 19 models): Model Score Rank (overall) GPT-6.1 Sol (OpenAI) 76.32 #1 Darwin-180B-RSI-R3 (VIDRAFT) 73.68 #2 Gemini 4 Argon (Google) 71.05 — DeepSeek V4.1 Flash 71.05 — Grok 4.7 (xAI) 68.42 — Qwen3.8-Flash (base, Alibaba) 57.89 — Darwin is #1 among all open-weight models in this category. SuperCLUE Overall (5 domains combined): Darwin: 43.08 (9th overall, 5th among open-weight models) Base Qwen3.8-Flash: 39.81 → RSI improvement: +3.27 points Other public benchmarks: GPQA Diamond (PhD-level physics, chemistry, biology): Darwin-family model scored 94.44 on the Hugging Face GPQA Diamond evaluation (published November 3, 2026) MMLU-Pro: Performance improved versus the base model (exact delta not disclosed publicly) Reasoning chain length: Reduced by 11% compared to base, meaning Darwin reaches correct answers with shorter chain-of-thought traces — a meaningful efficiency signal for production deployments Caveats: SuperCLUE evaluates primarily Chinese-language LLMs. Darwin's overall rank (9th) reflects weaker performance in long-context processing and instruction-following domains relative to its math/science strength. Engineers building multilingual or instruction-heavy applications should factor this in. How to try it Darwin model weights are publicly available on Hugging Face. You can browse and download the Darwin family (including POCKET-Darwin-180B for CPU inference) directly: # Browse available Darwin checkpoints huggingface-cli search vidraft # Download a specific Darwin model (replace MODEL_NAME with the actual repo ID from the HF page) huggingface-cli download vidraft/MODEL_NAME ⚠️ Substitute the actual repository name from VIDRAFT's Hugging Face page — do not guess model repo slugs. The POCKET-Darwin-180B variant is specifically designed for local CPU deployment on consumer hardware if you want to run inference without a GPU. Cumulative Hugging Face downloads across the Darwin model family have exceeded 4.68 million as of this report. FAQ Q: What makes RSI different from standard RLHF or DPO fine-tuning? A: Standard RLHF and DPO require human-annotated preference data or curated answer datasets at each improvement cycle. RSI as implemented in Darwin is self-contained — the model generates its own problems, produces candidate solutions, and verifies them autonomously. Human labelers are not in the loop for each iteration, which makes the improvement process more scalable and cheaper to run across multiple rounds. Q: Can I run Darwin on a machine without a GPU? A: Yes — POCKET-Darwin-180B is a compressed variant specifically optimized for CPU-only inference on personal PCs and laptops. For full 180B inference with the standard MoE model, GPU resources will still be required, though the sparse activation (~3B active params) reduces per-token compute costs significantly compared to a dense 180B model. Q: Darwin scores #2 in math/science but #9 overall on SuperCLUE — what's the gap? A: The RSI training is targeted specifically at reasoning-related weights. As a result, long-context processing and precise instruction-following — two of SuperCLUE's five evaluation domains — currently lag behind frontier proprietary models. VIDRAFT's public statement acknowledges this directly. If your use case is math, science, or logical reasoning, Darwin is competitive at the top of the open-weight leaderboard; for general-purpose assistant tasks, benchmark your specific workload before committing. Originally reported by AI타임스 (2026-11-06) — source article.