Z.ai releases open-weight GLM-5.3, tops agentic coding and cyber benchmarks

GLM-5.3, built on the same base as GLM-5.2 with extended post-training, scores 84.5% on CyberGym and 60 on the Artificial Analysis Intelligence Index, tying Kimi K3. Weights are on Hugging Face; running FP8 needs 10–12× H100s.
How this story unfolded
3 weeks · 13 reports · 21 community posts · 34 of 37 shown
- Aug 14
Z.ai Ships GLM-5.3 Without Retraining the Base Model: Better at Complex Coding and Long-Horizon Tasksmarktechpost.com
GLM-5.3 didn’t change the base model — where did its coding gains come from?thenewstack.io
China's Z.AI Ships GLM-5.3, Calling It the Top Open-Weight Coding Modeldecrypt.co
GLM-5.3: How Chinese labs keep stride with the frontierinterconnects.ai
GLM-5.3 is here with advanced cyber capabilities — and reportedly already found a 'serious vulnerability' in Cursorventurebeat.com
- Aug 15
- Aug 16
- Aug 17
- Aug 18
- Aug 19
GLM-5.3 hits the API at $1.4/$4.4 per million tokensventurebeat.com
An industrial-scale distillation of models, or subtle benchmaxxing: What developers really think of GLM-5.3thenewstack.io
GLM-5.3 achieves 60 on the Artificial Analysis Intelligence Index, on par with Kimi K3 and up 7 points from GLM-5.2. Once the weights are released it will be tied as the leading open weights model
- Aug 28
- Aug 29
- Aug 31
- Sep 1
More stories today
Hasan Toor: GPT-4o, Gemini 2.x don't watch video end-to-end
Hasan Toor·3 hours ago
Elon Musk says AI will be in charge, not humans
In a short clip from Tesla's 2025 Annual Shareholder Meeting, Elon Musk states that AI will be in charge, not humans.
YouTube·3 hours ago
Instagram's AI detection misfires, flags original photos
Users report Meta's "AI Content" label is being applied to original photos edited with tools like Canva's Background Remover, while actual AI imagery goes untagged. A similar issue occurred in 2024 with the "Made by AI" label.
The Verge·3 hours ago

Z.ai by email
Get an email when Z.ai has news
No news that day, no email.
Coder launches Agent Relay with SpaceXAI to run coding agents on own infra
Coder announced Coder Agent Relay, with SpaceXAI as launch partner, letting engineering teams run coding-agent tools on their own infrastructure while Cursor handles inference and planning in the cloud. One Coder engineer noted 1% of engineers responsible for 40% of token spend.
The New Stack·3 hours ago

Cognition's math Olympiad gold medalists get token budgets for math problems
Emad Mostaque·3 hours agoFL Studio CEO says GenAI can bring beginners into music-making
Image-Line's CTO Robert Linke says generative-AI tools can funnel complete beginners into human music-making, leading them to download FL Studio. He stressed AI features must be ethically trained and better than traditional solutions.
Music Ally·3 hours ago

AI coding agents install untrusted code from poisoned llms.txt files
Researchers scanned 6,214 corporate domains, finding 120 llms.txt files pointing to unregistered packages. After registering some, they got phone-home responses from Fortune 500 companies within an hour, implicating Claude, OpenAI's Codex, and Hermes.
Schneier on Security·4 hours ago

AI job applications create an infinite doom loop
Job seekers use AI to craft applications while employers use AI to screen them, but automated ranking isn't always used, and many applicants hear nothing back. One candidate, Christopher, applied to 700 jobs and had ChatGPT talk to an AI recruiter for 10 minutes.
Wired·5 hours ago
