Darwin-180B-RSI: VIDRAFT's Open-Weight MoE Reasoning Model Hits #2 on SuperCLUE and #1 on GPQA Diamond
TL;DR: VIDRAFT's open-weight reasoning model, Darwin-180B-RSI, scored 73.68 on the SuperCLUE November math & science reasoning leaderboard—ranking 2nd globally across all models—and topped the Hugging Face GPQA Diamond leaderboard with 94.44. Built on a 180B-parameter MoE architecture with Recursive Self-Improvement (RSI), it's fully open-weight, downloadable from Hugging Face, and comes with a compressed on-device variant called Pocket Darwin 180B.
What it is
Darwin-180B-RSI is VIDRAFT's open-weight large language model specialized in complex mathematical and scientific reasoning. Key facts from the announcement:
- Architecture: 180 billion total parameters in a Mixture-of-Experts (MoE) structure, with only ~3 billion parameters activated per forward pass during inference.
- Base model: Built on top of Alibaba's Qwen3.8-FlashNext as the foundation.
- Training approach: VIDRAFT applied their own model-level Recursive Self-Improvement (RSI) methodology on top of the base model.
- Weight availability: Fully open-weight — available for research and commercial use.
- On-device variant: A compressed version called Pocket Darwin 180B is provided for use on personal PCs and laptops without a dedicated GPU.
- Cumulative downloads: Over 4.68 million downloads on Hugging Face as of the announcement date.
- Leaderboard distinction: The only Korean model in the SuperCLUE November 2026 evaluation.
How it works
At a conceptual level, Darwin-180B-RSI is built around the idea that a model can improve its own reasoning capability without requiring additional human-labeled ground-truth data.
VIDRAFT's RSI approach works like this:
- Base model selection: Start with a capable open-weight MoE base model (Qwen3.8-FlashNext in this case).
- Selective weight tuning: The base model's expert feed-forward network (FFN) layers are kept frozen. VIDRAFT's RSI process adjusts the weights of the attention layers, linear attention layers, and shared expert layers.
- Self-generated training signal: Rather than providing curated human-annotated datasets, the model itself generates problems, produces solutions, and iteratively verifies those solutions. This closed-loop cycle is what VIDRAFT calls Recursive Self-Improvement.
- Reasoning specialization: Because the RSI loop is tightly focused on problem generation and verification, the resulting model becomes particularly strong at structured reasoning tasks — mathematics, physics, chemistry, and biology — rather than general-purpose generation.
This approach is architecturally distinct from standard supervised fine-tuning or RLHF pipelines in that the improvement signal is internally generated rather than externally provided.
Benchmarks & results
All numbers below come directly from the source article.
SuperCLUE November 2026 Comprehensive Evaluation (Chinese LLM evaluation body; uses private held-out test items across 5 domains: agentic coding, math/science reasoning, hallucination suppression, instruction following, and long-context handling):
| Model | Math & Science Reasoning | Overall Score | Overall Rank |
|---|---|---|---|
| OpenAI GPT-6.1-Sol | 76.32 | — | #1 |
| Darwin-180B-RSI | 73.68 | 43.08 | #2 (math/sci) / #9 overall |
| Google Gemini 4-Argon | 71.05 | — | — |
| DeepSeek V4.1 Flash | 71.05 | — | — |
| Grok 4.7 | 68.42 | — | — |
| GLM-5.3 | 65.79 | — | — |
| Kimi K3.1 Pro | 63.16 | — | — |
| Qwen3.8-FlashNext (base) | 57.89 | 39.81 | #15 |
- Darwin-180B-RSI ranks #2 overall in math & science reasoning across all 19 models evaluated.
- Among open-weight models only, Darwin ranks #5 overall and #1 in math & science reasoning.
- The base model (Qwen3.8-FlashNext) scored 57.89 on the same math/science sub-task — Darwin's RSI training delivered a +15.79 point improvement on that dimension.
Hugging Face GPQA Diamond Leaderboard (PhD-level physics, chemistry, and biology questions; results published November 3, 2026):
- Darwin-180B-RSI scored 94.44, ranking #1 overall.
How to try it
Darwin-180B-RSI is fully open-weight and publicly available on Hugging Face. With over 4.68 million cumulative downloads, the model is actively used by the community.
You can browse and download the model using the Hugging Face CLI:
pip install huggingface_hub
huggingface-cli download VIDRAFT/Darwin-180B-RSI
Note: The exact repository slug above follows standard Hugging Face naming conventions — confirm the exact repo name on the VIDRAFT Hugging Face page before running.
The compressed Pocket Darwin 180B variant is also available on Hugging Face for on-device inference without a GPU — suitable for local experimentation on consumer hardware.
No API endpoint details are publicly disclosed in the source article.
FAQ
Q: How does Darwin-180B-RSI compare to its base model, Qwen3.8-FlashNext?
A: On the SuperCLUE math & science reasoning sub-benchmark, Darwin-180B-RSI scored 73.68 versus the base model's 57.89 — a gap of nearly 16 points. On the overall SuperCLUE ranking, Darwin moved from 15th (base model) to 9th, using the same expert FFN layers but with RSI-adjusted attention and shared expert weights.
Q: Can I run this on consumer hardware?
A: Yes. VIDRAFT publishes Pocket Darwin 180B, a compressed variant of the model specifically designed to run on personal PCs and laptops without a dedicated GPU. The full 180B MoE model activates only ~3B parameters per token, which helps with inference efficiency, but the Pocket variant goes further with model compression for truly on-device use.
Q: Is this model suitable for commercial use?
A: The source article states the model is released as fully open-weight and available for both research and commercial use. Check the specific license file in the Hugging Face repository for the exact terms before deploying in production.
Originally reported by 디지털데일리 (2026-11-06) — source article.
Top comments (0)