VIDRAFT, November 1 to 8, 2026: Open-Weight #1 in Math and Science Reasoning, and a Gate That Stopped 290 of 290
aillmbenchmarkssecurity
What did VIDRAFT ship this week?
Three results, each checkable on a public page.
- SuperCLUE, November 2026. Darwin-180B-RSI-R3 scored 73.68 in mathematical and scientific reasoning: first among open-weight models and second of all 19 models, behind only OpenAI's GPT-6.1-Sol at 76.32 and ahead of Google Gemini-4-Argon, DeepSeek-V4.1-Flash, xAI Grok-4.7 and Qwen3.8-Max. It is the only Korean model on the board. Scores are SuperCLUE's, measured on its private November test set.
- ViKernel. A backdoored model passed a 24-check scan and then emitted attack commands. The gate denied 290 of 290 payloads at a 237 ms median, and 661 exfiltration probes produced 0 leaks.
- POCKET. The on-device builds passed 3.15 million downloads, with the 180 billion parameter build running from SSD on a laptop.
As of November 8, 2026.
How far did Darwin move past the model it inherited from?
Darwin-180B-RSI kept the 512 routed experts per layer from Alibaba's Qwen3.8-Flash-Next byte for byte and rebuilt only the attention layers, the linear attention layers and the shared experts through recursive self-improvement. Both models sit on the same board.
| Qwen3.8-Flash | Darwin-180B-RSI-R3 | |
|---|---|---|
| Math and science reasoning | 57.89 | 73.68 |
| Overall intelligence index | 39.81 (15th) | 43.08 (9th) |
Same experts, same stored knowledge, rebuilt reasoning layers: 15.79 points on the hardest axis and six places on the overall index.
Two days before the SuperCLUE board, the same model took first place overall on the Hugging Face GPQA Diamond leaderboard at 94.44, a PhD-level science test in English.
Who reported it?
Eighteen outlets in seven countries during the week, all listed with links on the VIDRAFT press page.
| Country | Outlets |
|---|---|
| Korea | Chosun Biz, Hankyung, Hankook Ilbo, Economist Korea, Digital Daily, AI Times, ZDNet Korea, Business Report |
| China | NetEase (four articles), AI Pulse |
| Japan | Labmemo |
| Germany | Mindverse |
| Argentina | Blog Donweb |
| Global English | llms.blog, ByteWoops, Night Grade, LUCAS AI News |
What can you open right now?
- Agent Guard, six live scenarios: ViKernel Desktop. No login, no key.
- Weights: Darwin-180B-RSI-R3, open weights under the Qwen community licence, commercial use permitted.
- On-device build: POCKET-Darwin-180B-GGUF, 4-bit, runs from SSD on a 32 GB laptop with no datacenter GPU.
- Code: vikernel-agent-guard.
Our pages on each result
- A model scan found nothing. The gate stopped 290 of 290 attack commands
- Darwin-180B-RSI ranks #1 open-weight model in math and science reasoning on SuperCLUE
- Korean press on Darwin's SuperCLUE result
- An AI agent you can stop: the kill switch lives inside the model
- POCKET passes 3 million downloads with no GPU
- VKORE-325M: pretraining a language model on an inference-only NPU
- Top-300 open LLM downloads: how the Qwen share moved
FAQ
Is Darwin-180B-RSI-R3 first overall on SuperCLUE? It is first among open-weight models and second of 19 in mathematical and scientific reasoning. On the five-axis overall index it places 9th of 19 and 5th among open-weight models.
Who produced the SuperCLUE scores? SuperCLUE, on its private November 2026 test set. VIDRAFT is an evaluated model on that board, not the operator of the benchmark. The board is at superclueai.com.
What is recursive self-improvement here? The model writes problems, solves them, verifies the answers, and updates its own weights from the verified traces. In Darwin-180B-RSI it rebuilt the attention and shared-expert layers while the inherited expert layers stayed byte-identical.
What licence covers the weights? Qwen community licence 1.0, inherited from the model whose expert layers Darwin reuses. Commercial use is permitted.
Which numbers here are VIDRAFT's own measurements? The ViKernel gate figures: 290 of 290 denied, 237 ms median, 661 probes with 0 leaks, and the live verdicts in the linked article. The SuperCLUE scores are SuperCLUE's, and the GPQA Diamond rank is the Hugging Face leaderboard's.