Darwin-180B-RSI-R3: VIDRAFT Releases Weights for Its Recursively Self-Improved 180B MoE Reasoning Model

Darwin-180B-RSI-R3: VIDRAFT Releases Weights for Its Recursively Self-Improved 180B MoE Reasoning Model TL;DR: Korean AI startup VIDRAFT has open-released the weights for Darwin-180B-RSI-R3, a 180B-parameter Mixture-of-Experts model trained through iterative Recursive Self-Improvement (RSI). On the SuperCLUE November 2026 math and science reasoning evaluation, it scores 73.68 — first place among all open-weight models and second overall. If you're building reasoning-heavy applications and need a deployable, permissively licensed frontier-class model, this is worth your attention. What it is Darwin-180B-RSI-R3 is a large language model developed by VIDRAFT, a Korean Pre-AGI AI startup. Key facts from the public model card: Architecture: Mixture-of-Experts (MoE) with ~180B total parameters; approximately 3B parameters are activated per forward pass during inference. Base model: Built on Alibaba's Qwen3.8-Flash-Next as the starting checkpoint. Training paradigm: Recursive Self-Improvement (RSI) — the "R3" suffix denotes the third iteration of this self-improvement loop. Weight availability: Full weights are publicly released on Hugging Face. A 4-bit GGUF quantization (~111 GB) is also available, making CPU-only inference on a high-RAM workstation feasible without a GPU. License: Weights are publicly released (check the Hugging Face model card for the exact license terms before production use). The model is focused on mathematical and scientific reasoning — it is not positioned as an all-around general assistant, and the source reporting explicitly notes gaps in long-context handling and instruction following relative to closed competitors. How it works The core idea behind RSI is that the model bootstraps its own training signal rather than relying on human-annotated solutions or externally provided reasoning traces. At a conceptual level, the loop works like this: Problem generation / selection: The model is given (or selects) problems that have verifiable correct answers — typically formal math or science problems where correctness can be checked programmatically. Self-solving: The model attempts solutions autonomously. Verification: Each candidate solution is checked against the verifiable ground truth. Only solutions the model itself got correct are retained. Data recycling: Those verified, self-generated correct solutions become new fine-tuning data for the next iteration of the model. Repeat: The improved model becomes the starting point for the next RSI round (hence R1 → R2 → R3). Crucially, no human-written solutions and no human reasoning traces are used in the RSI fine-tuning loop. The improvement signal is entirely derived from the model's own verifiably correct outputs. This is a data-efficient approach to capability improvement that sidesteps the bottleneck of expert annotation at scale. The parent model (Qwen3.8-Flash, before VIDRAFT's RSI treatment) scored 57.89 on the same SuperCLUE math/science benchmark. After three RSI rounds, Darwin-180B-RSI-R3 reaches 73.68 — a substantial lift attributable to the iterative self-improvement process rather than a change in model scale or architecture. Benchmarks & results All numbers below are from VIDRAFT's public model card and the SuperCLUE November 2026 evaluation (non-public test set, scores as reported by VIDRAFT): Model Developer Open Weights Math & Science Reasoning GPT-6.1-Sol OpenAI ❌ 76.32 Darwin-180B-RSI-R3 VIDRAFT ✅ 73.68 Gemini-4-Argon Google ❌ 71.05 DeepSeek-V4.1-Flash DeepSeek ✅ 71.05 Grok-4.7 xAI ❌ 68.42 Qwen3.8-Flash (base) Alibaba ✅ 57.89 Darwin-180B-RSI-R3 is ranked 2nd overall (out of 19 models) and 1st among all open-weight models in math and science reasoning. It is the only Korean model to appear in the SuperCLUE November 2026 rankings. On the 5-domain composite index, the model scores 43.08, placing 9th overall and 5th among open-weight models — the source explicitly flags long-context and instruction-following as relative weaknesses. How to try it The weights are publicly available on Hugging Face. You can download the model or the quantized GGUF variant using the Hugging Face CLI: pip install huggingface_hub huggingface-cli download VIDRAFT/Darwin-180B-RSI-R3 For the 4-bit GGUF version (suitable for CPU inference on a machine with sufficient RAM): huggingface-cli download VIDRAFT/Darwin-180B-RSI-R3-GGUF ⚠️ Verify the exact repository slug and filenames on the official Hugging Face model card before running — do not rely on the above paths as canonical. Check the license on the model card before any production or commercial deployment. FAQ Q: Why does the composite score (43.08, 9th overall) look so much weaker than the math/science score (73.68, 2nd overall)? A: Darwin-180B-RSI-R3 was explicitly optimized for mathematical and scientific reasoning through the RSI loop. The composite index aggregates five domains — including long-context understanding and instruction following — where the model was not specifically trained to improve. Specialization is a deliberate trade-off here, not an oversight. Q: Can I actually run a 180B MoE model on consumer hardware? A: The 4-bit GGUF quantization weighs in at approximately 111 GB. That rules out most consumer GPUs, but it does open the door to high-RAM CPU workstations or multi-GPU setups. With MoE only activating ~3B parameters per token, inference throughput is meaningfully better than a dense 180B model at the same quantization level. Expect slow generation on CPU-only hardware — it's feasible for experimentation, not production throughput. Q: Does this model replace general-purpose LLMs for my application? A: Probably not if your use case involves long documents, complex multi-turn instructions, or domains outside math and science. Darwin-180B-RSI-R3 is a strong choice for tasks with verifiable, structured reasoning — theorem proving, competition math, scientific QA — where its benchmark lead is most relevant. Originally reported by 라보메모 ラボメモ (일본) (2026-11-07) — source article.