🚀 Zhihu Frontier is now live. We're here to bring China's AI & tech trends, voices, and perspectives to the global stage — powered by Zhihu Inc., China's leading knowledge platform. 🧭 Exploring the frontiers of technology and ideas. 📌 From developer insights to deep tech

📢 ICML 2025 | From Language to Vision: VARSR unlocks a new paradigm for Image Super-Resolution. Recent advances show autoregressive models (AR) - successful in NLP - are now thriving in vision tasks (like DALL·E, GPT-4o). Compared to Diffusion Models, AR methods better capture

🧠 Zhihu contributor & @Kimi_Moonshot senior researcher @Jianlin_S shares insights on K2’s architecture & optimization:.🔥One key breakthrough? Solving the MaxLogit explosion with a combo of Muon + QK-Clip.(Technical details see previous post👇).🧩 More reflections on model.

🧠Zhihu contributor & @Kimi_Moonshot dev Dylan shares his thoughts on building Kimi K2:. Why RL? Because compute may be infinite, but data is not. RL improves data efficiency - that's why we invest in scaling test-time compute. Why large models? Why Muon optimizer?.→ It's all

🤖 Zhihu contributor & @Kimi_Moonshot RL Lead @RotekSong shares how they pushed Kimi K2 toward better general-purpose Agent abilities - by scaling up tool-use data. In short: a fully automated agent data factory 🏭 that simulates end-to-end workflows to filter out high-quality

🚀 Zhihu contributor & @Kimi_Moonshot senior researcher @Jianlin_S drops a new piece:."QK-Clip: Taking Muon Further on the Scaleup Journey". While scaling Muon to 100B+ params, a new bottleneck hit: MaxLogit explosion 💥.Enter QK-Clip - a post-hoc fix to Q/K weights (unlike

🎉 Great to see @Kimi_Moonshot infra dev 刘少伟 sharing why Kimi K2's config "looks the way it does" - from an inference perspective 🤖. 🧩 Constraints:.Inherits DSv3 structure.Adjust internal model parameters to fit needs.Training & inference cost ≈ DSv3.🎯 Goal: Lower loss.

🚀🔥 @Kimi_Moonshot drops the K2 model — now #1 trending on Hugging Face!.💡 Devs break it down, tech folks dive deep. 👀📎 Full discussion & hands-on reviews on Zhihu:.#Kimi #K2Model #LLM #HuggingFace #AI.

🚀 Kimi releases its first trillion-parameter open-source agentic model — K2!.What's new? What's powerful? What's next? 🤖🔥.Full discussions, hands-on reviews & benchmarks now live on Zhihu:.👉 Welcome to join the convo! 🧠💬.Try it now at.

🧠 Zhihu contributor @langfengq shared his thinkings about open-sourced verl-agent - an RL framework designed to train reasoning-capable LLM agents!. It extends veRL, and unlike methods that simply concatenate full interaction history, verl-agent treats each step as an

💬 How do people view Grok 4, the new-gen model from Musk’s xAI? What are its standout features?. 🔗Dive into discussions from Chinese tech community Zhihu users here:.#Grok4 #xAI #AI.

Zhihu contributor 周国睿 has spent the past year exploring a big Q:.👉 Can we take recommender systems to the next level?. Based on hands-on work with OneRec, they found E2E recsys might be the future. Here are his key thoughts:.🔍How big should a rec model be? .⚙️How to scale

🧠 Zhihu contributor @jackbai_jkb shares insights from his paper "Thinking vs. Doing: Agents that Reason by Scaling Test-Time Interaction". How to do multi-step training on a model post-trained in single-step environments? They found the ideal training setup should be:.⚙️ no

「Tech Insight」.🔥 #HackerNews Top Post: "The new skill in AI is not prompting, it's context engineering" — thoughts?. 💬 Zhihu contributor Navis Li:.What really makes or breaks your AI app is what kind of "ingredients" you feed it. Context Engineering = crafting the entire

🚀 What's the hype about @Zai_org open-source GLM‑4.1V‑9B‑Thinking model?. 🧑‍💻 Zhihu contributor 社恐打工仔:.Pretty slick! Fast inference⚡. Just 9B params yet beats many bigger models. Why? It thinks, not just predicts. 🔥. 🧠 Zhihu contributor Mirabella:.Breakthrough design -

🎙️ AI Insight Talk | Code Bench Special Live!. 🗓️ July 3rd, 8:30–9:30 PM (CST).🔗 Co-hosted by @huggingface × @OpenMMLab × @MaaSAI42 × Zhihu × 机智流 & more!. Topics:.🔍 CPRet – Is the model "memorizing" or truly understanding? Exposing performance inflation on similar problems

「Developer Talk」.🧠 In the RL Scaling era, what kind of RL framework do we need?.🚀 To answer that, THUDM open-sourced their self-developed framework slime: 👨‍💻 RL infra @Zai_org & Zhihu contributor @朱小霖 breaks it down:.🧩 Believes LLM + RL = final

🔥 Baidu open-sourced the ERNIE 4.5 series!.🤖 What does it mean for China's AI ecosystem?.📊 Reviews & Discussions on Zhihu:.#AI #LLM #OpenSource #MoE.

冯一尘 mentioned two important things for building an Agent:.1️⃣ Agents that can think long 🤔.2️⃣ End-to-end Reinforcement Learning 💡.Another Kimi’s dev Flood Sung explains why "long thinking" matters, and bonus discovery during training:.📈As performance improves, token usage.

「AI Frontier」.🧪 @Kimi_Moonshot 's first Agent - Kimi-Researcher - entered limited rollout on June 20. It generates long-form, source-traceable research reports. 👨‍💻 Kimi engineer 冯一尘 personally responded on Zhihu:.We're not building a search tool. We're training an AI Agent

「Tech Insight」.🧠 Zhihu contributor 许华哲Harry shares his take on Embodied Intelligence:. He reflects on common failure modes in this field:.• Chasing the "coolest task" at all costs.• Building virtual worlds and hoping digital solves everything.• Dumping massive data into