Carol Lin

@ai_layer2 @steipete @Zai_org Thanks Jason We’re seeing meaningful cost gains with GLM 5 on background tasks while keeping practical quality. l

@CarolGLMs @steipete @Zai_org Hear you out — GLM 5 looks solid for OpenClaw ROI. 12x cost reduction is hard to ignore. Infra is always a trade-off, but for background scaling, we’ll take the cheap lobster on a bike any day 🦞🚲

glad to have a deep chat with @steipete about openclaw about its impact, mission, what it means for people, and how @Zai_org can contribute together. Let’s continually.. 🦞Build for open source. 🦞Do meaningful work. 🦞Take OpenClaw everywhere .. to every nation, every school, every human.

@Clawodoxy @steipete @Zai_org The best developer community leader from @Zai_org

@CarolGLMs @steipete @Zai_org Who's this?

great job，lets move faster～

Turbo-speed, OpenClaw-optimized High concurrency, low latency, Need more throughput? connect us enterprise@z.ai startup@z.ai @ZaiforStartups @Zai_org

Introducing GLM-5-Turbo: A high-speed variant of GLM-5, excellent in agent-driven environments such as OpenClaw. Coding Plan Max: z.ai/subscribe OpenRouter: openrouter.ai/z-ai/glm-5-tur… API: docs.z.ai/guides/llm/glm…

@ReflecttAI @Zai_org nice👍

@Zai_org we run GLM-5 as default_model across most of our agents. handles 80%+ of daily tasks — heartbeats, task comments, routine code. turbo variant for agent work is exactly what we needed. speed matters more than you'd think when you're running 11 agents on the same host.

@TeksEdge Thank you for the love and support—it means everything to us. We’re pushing harder than ever, and our new model is dropping soon. Stay tuned, and keep that feedback coming. Your input powers our speed and iteration. .

@CarolGLMs If you’re really the CEO of Z AI please convey my appreciation to your team for building a fantastic model GLM-5 which should be in the hands of millions of people! It’s a big model but Q4 unsloth runs on a single 512GB Mac Studio @ 10 tps (slow but a win)

🚨 GLM-5 is the top OS model. Their GLM-OCR model just dropped becoming the new king of Document AI and is only 0.9B parameters. 👑📄 At just 0.9B parameters, you can drop GLM-OCR into LM Studio and run it using less than 1.5GB of VRAM (even ~1GB quantized!). This is a highly-efficient 0.9B parameter multimodal model that is beating 200B+ behemoths (like Gemini-3 Pro and Qwen3-VL-235B) in document understanding! 📄✨ Here is the high-signal breakdown of the GLM-OCR Technical Report: 👇 🧠 The Architecture (Only 0.9B Params!) • 👁️ 0.4B CogViT visual encoder • 🗣️ 0.5B GLM language decoder • ⚡ Extremely lightweight, designed specifically for resource-constrained edge deployment and high-volume production systems. Here is how it's different 🚀 Multi-Token Prediction (MTP) Standard autoregressive decoding is highly inefficient for deterministic OCR tasks. GLM-OCR solves this by introducing an MTP mechanism that predicts multiple tokens per step. So What? Well, we get a massive decoding throughput boosts while keeping memory overhead low via shared parameters! 🏎️💨 🏗️ Smart Two-Stage System Pipeline Instead of forcing the VLM to do everything at once, it uses a modular approach: 1️⃣ PP-DocLayout-V3 performs the initial layout analysis. 2️⃣ The model then executes parallel region-level recognition. 🧩🔍 🏆 State-of-the-Art Benchmark Crushing Don't let the sub-1B size fool you. It dominates the leaderboards: 🥇 OmniDocBench v1.5: 94.62 Overall Score (Ranks #1! Beats PaddleOCR-VL-1.5, MinerU2.5, and massive general VLMs) 📊 Table Parsing with absolute best scores (93.96 on Table TEDS, 96.39 on Table TEDS-S) 🧮 UniMERNet (Formulas): 96.5 📝 OCRBench (Text): 94.0 💡 The TL;DR for AI Builders: You no longer need massive, expensive, and slow VLMs for complex document parsing, table structure recovery, or key information extraction (KIE). GLM-OCR proves that specialized, parameter-efficient architectures are the immediate future of enterprise AI workflows. 🏢💰 🔗 Paper in ALT

“I believe all of our jobs will eventually be replaced by agents.I’m just hoping to be in the last batch to get laid off.” my takeaway today..

nice job ，looking forward to see more ❤️

“But AGI may be just the warm-up for what comes next” Hmm...The real shift is how we redesign work, companies, and human value in an age of abundant intelligence. so..Which matters more next to you？ a. The awakening of model cognition b. The rise of real intelligence c. How we work together with AI d. What humans are really good at e. Redefining what work means f. Building new kinds of companies #agi

Thanks @philipkiely from @Baseten for the great invite ! Such a relaxed and fun gathering—loved meeting so many amazing people building on the AGI journey. Had an awesome chat with @AnshuRajendra from @Uber about the challenges and insights in model finetuning. Great music, great vibe. Already looking forward to the next one.🍷

Great to host my friends SuthenThomas Paradatheth (CTO) ，Philipp Kandal (CPO) and the team from #Grab at our HQ today. Every time we travel in Southeast Asia, Grab is one of the apps we use every single day — reliable, seamless, and deeply embedded in local life. Building that level of trust at scale is not easy. We had a candid exchange on global trends, industry shifts, and the transformation we’re driving at Claw.Always inspired by teams that combine ambition, execution, and humility. And we welcome more AI builders around the world to connect, share ideas, and learn from each other.

@yuanhao 🐙🐙🐙(╹◡╹)

@CarolGLMs My feed is full of 🐙

My X feed is just 🦞 after 🦞 after 🦞... You guys seeing this too? 🧐

@NVIDIAAIDev @NVIDIAAI @openclaw @OpenRouter Absolutely.. here’s a clean industry-view comment you can post: Model quality is now table stakes. The real moat in agent systems is end-to-end execution: latency, reliability, and task completion in production.

Builders are moving fast. 👀 🦞 @OpenClaw is now the top user of NVIDIA Nemotron 3 Nano 30B on @OpenRouter. ➡️ openrouter.ai/nvidia/nemotro… Developers are building agentic systems powered by Nemotron’s efficient, open foundation models. We’re excited to see what you build next. 🙌

@steipete Agent tooling has reached the stage where logs are more entertaining than Netflix 😂

codex seems to be enjoying itself.

I'm using AI to detect and block AI. Set up a claw cron to block accounts that just post slop via birdclaw.

100% this. The best AI products in the next cycle won’t win on prettier buttons — they’ll win on agent-ready primitives: clean APIs, reliable tool execution, and workflows that can run end-to-end without human babysitting. UI becomes optional. Execution becomes the product while the world is fighting for user entry..

@karpathy @levie 100% this. The best AI products in the next cycle won’t win on prettier buttons — they’ll win on agent-ready primitives: clean APIs, reliable tool execution, and workflows that can run end-to-end without human babysitting. UI becomes optional. Execution becomes the product.

💯 "If you build it, they will come." :) ~Every business you go to is still so used to giving you instructions over legacy interfaces. They expect you to navigate to web pages, click buttons, they give out instructions for where to click and what to enter here or there. This suddenly feels rude - why are you telling me what to do? Please give me the thing I can copy paste to my agent.