MetaStone-S1 just dropped on Hugging Face Test-Time Scaling with Reflective Generative Model Experiments demonstrate that MetaStone-S1 achieves comparable performance to OpenAI-o3-mini's series with only 32B parameter size

Algorithms and data always move forward in an iterative way.

Say hello to MetaStone-S1！

human agent based on chatgpt🔍

7 notable AI models to pay attention to: ▪️ Kimi K2 ▪️ Gemini 2.5 ▪️ Grok 4 ▪️ Voxtral ▪️ MirageLSD ▪️ LTX-Video ▪️ MetaStone-S1 🧵

XBai o4 has released the core technology of Gemini 2.5 Deep Think in “parallel thinking”🔥Google DeepMind

Thanks Ivan, anyone can enjoy our XBai o4 in this way

Big move

🚀 Just published: “XBai o4: A Guide With Demo Project” Walks you through locally deploying MetaStoneAI's new reasoning model with LM Studio and Streamlit . Dive in🧵

🎉XBai o4 is now live — Experience the power of parallel think at chatxbai.com ! • Smarter web search — 93.2 SimpleQA 🔍 • Outperforms OpenAI’s o3-mini-medium in reasoning 📈 • Fast inference at 50 tokens/s ⚡️ API: Email [email protected] with a short

We’ve come up with a new way to do parallel thinking in interactions — hope everyone enjoys it! By the way, it’s super fast and really useful.

Hi everyone, our latest graph-based synthetic data method is accepted by NIPS2025, which achieves 255 expanison ratio (e.g. use 1k seed data and get 255k high-quality new data). April version: arxiv.org/pdf/2412.08864 (camera ready will come in few days) Github: