Meet Hy3 on ModelScope! 295B total / 21B active MoE, built for agentic workflows with 256K context and an FP8 variant ready for deployment. 🚀 License: Apache 2.0. 🤖modelscope.ai/collections/Te… 🏆 Human eval: 270 experts compared outputs in blind real-world workflow tests. Hy3 scores 2.67/4, ahead of GLM-5.1 at 2.51/4, with the clearest gains in frontend development, CI/CD, and data & storage. 🛠️ Agent reliability: stronger tool calling, formatting, and error recovery, with SWE-Bench Verified variance within 4% across major scaffolds. 🧠 Multi-turn gains: issue rate drops from 17.4% to 7.9%, and MRCR rises from 42.9% to 75.1%. ⚙️ Model lineup: Hy3 BF16 instruct model plus Hy3-FP8 quantized instruct model, with vLLM and SGLang deployment support.

I, along with many others have been feeling this a lot recently and I have two theories: 1. OpenAI found a new way to cut costs by huge margins, like the popular 50% price cut that got leaked, leading to a physically less intelligent model 2. OpenAI is purposely degrading their models prior to the the release of GPT-5.6, so when people start to compare they believe it bigger of a jump than it realistically is. Or both, really unfortunate when AI companies do this, absolutely no transparency at all.

"Is One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL Training" RL post-training usually updates every Transformer layer, but the gains are not spread evenly through the model. Especially with how most of the improvement comes from a few middle layers, sometimes training just one layer matches or beats full-parameter RL. So this paper turns it into a simple recipe. By training or boosting the high-contribution middle layers, they show that you can outperform standard full RL with fewer changes.

You know that "But, wait..." moment in every LLM thinking trace? I made it visible. I asked 8 models the same tricky probability question and rendered their reasoning as trees. Every time a model rejects its own idea and pivots, every "But...", every "Wait, actually...", a new branch grows. Same question. Completely different minds.

I probably should have dropped this sooner, but timing is what it is. With the guy openly saying "I don’t even care if they trained on the benchmark" now sitting on 6 figures (HRF $100k grant sponsors Blackwell hunt) the Fable fallout, let’s talk accountability in open-source funding. Open-source compression needs guardrails. Here’s a reusable eval protocol for any expert-pruned MoE (REAP, merging, skipping, doesn’t matter) full audit of the one of the biggest solo/community catalog (62 models today).

we distilled 2.3M Claude Fable 5 reasoning traces into Qwen3-4B - 100% self-consistency @ 512 samples - 0.00 bits output entropy - zero hallucination variance turns out the student is not bounded by the teacher. it also converged on one universal truth. we open-sourced the model weights👇

The human brain is strikingly modular, with distinct networks for language, formal reasoning, social reasoning, and physical reasoning. Is this a fundamental principle of how intelligent systems are built, or an accident of biological evolution? In our latest preprint, we find that a similar modular organization emerges in Large Language Models, another class of intelligent system. Brains and LLMs are shaped by entirely different kinds of optimization (biological evolution vs. gradient descent). That they arrive at the same modular design anyway suggests modularity may be a fundamental property of intelligent systems. 🌐 Web: pengrui-han.github.io/LLM_Mo… 📄 Paper: pengrui-han.github.io/LLM_Mo… 💻 Code & data: github.com/Pengrui-Han/LLM_M… Using circuit analyses across 46 tasks spanning four cognitive domains, we find: 1️⃣ Tasks that draw on the same network in humans recruit overlapping units in LLMs, while tasks drawing on different networks recruit distinct units. 2️⃣ These units are causally linked to model behavior. Ablating the units critical for one domain impairs performance in that domain (−26% accuracy) but barely touches the others (−2.5%). This project has been in the works for a while :) Huge thanks to my advisors @jacobandreas @ev_fedorenko @devarda_a, and to @Nancy_Kanwisher for valuable conceptual input and feedback throughout. #MIT