A 35B AI Beat GPT-5.6. Here’s the Catch.

Viral posts are claiming that Shanghai AI Lab's new open-weight model, Sage-A1.5, "outshines the behemoths"—going toe-to-toe with and beating trillion-parameter frontier giants like Kimi K3, GLM-5.3, and GPT-5.6 Sol. Even wilder: Sage holds only 35 billion parameters and activates just 3 billion parameters per word. Can a small model running locally on consumer hardware really outthink frontier frontier models? In this video, we open InternScience's technical report and verify the actual scorecard. While the science breakthroughs are genuine—Sage beats GPT-5.6 Sol 81.0 to 73.0 on Olympiad-level science and doubles it on PhD research subtasks—the viral posts leave out the other half of the table: it consistently trails the leaders on web search, software engineering, and tool use. We break down the MoE architecture, how its parallel subagent delegation works, the self-verification trick, and what happened when a real user ran it locally on an AMD laptop chip and hit a 30,000-token infinite thinking loop. ⏱️ Chapters: 0:00 Can a 35B Model Beat Frontier Giants? 1:25 The Scorecard: Science Wins vs Coding Losses 2:19 Inside the 3B-per-Word MoE Architecture 3:38 Parallel Helpers: The Subagent Mechanism 5:17 Self-Verification and the Teacher Model Catch 7:10 Local 24GB Setup and the 30,000-Token Loop 8:01 Final Verdict: When Is It Worth Downloading? 🔗 Sources & Documentation: • Technical Report (arXiv 2610.28578): https://arxiv.org/abs/2610.28578 • HTML Paper & Scorecard: https://arxiv.org/html/2610.28578v1 • Sage-A1.5 Model Card: https://huggingface.co/InternScience/Sage-A1.5 • Sage-A1.5 Architecture Config: https://huggingface.co/InternScience/Sage-A1.5/raw/main/config.json • Qwen3.6-35B-A3B Base Layout: https://huggingface.co/Qwen/Qwen3.6-35B-A3B • Kimi K3 Model Card Comparison: https://huggingface.co/moonshotai/Kimi-K3 • Official GGUF Local Downloads (Q4_K_M): https://huggingface.co/InternScience/Sage-A1.5-Q4_K_M-GGUF • Hugging Face Discussion (Ryzen 30k Token Loop): https://huggingface.co/InternScience/Sage-A1.5/discussions/2 • Hugging Face Discussion (Size Class Comparisons): https://huggingface.co/InternScience/Sage-A1.5/discussions/3 🔔 Subscribe: https://www.youtube.com/channel/UC0DZj1PNa_Fp0MU6uPSKv5w?sub_confirmation=1 💙 Become a Member: https://www.youtube.com/channel/UC0DZj1PNa_Fp0MU6uPSKv5w/join 🐦 Twitter/X: https://x.com/cloud_codes 💬 Discord: https://discord.gg/mhgXXmhrh #ai #machinelearning #opensourceai #llm #datascience #deeplearning #artificialintelligence #cloudcodes