Source: AI News — 2026-08-14
Summary
Zhipu released GLM-5.3, a 744-billion-parameter mixture-of-experts model post-trained from the GLM-5.2 base, with a 200K-token context window and an open-weight license. The headline numbers: its Terminal-Bench 3.0 score jumped from 4.6 to 28.3 — a 6.2x improvement — DeepSWE 1.1 hit 66.9%, and on CyberGym, a cybersecurity capability benchmark, it scored 84.5%, edging out Anthropic's Mythos 5 (83.8%) and OpenAI's GPT-5.6 Sol (83.6%).
Key Takeaways
- Architecture: 744B total parameters, mixture-of-experts, post-trained from the GLM-5.2 base, 200K-token context window, open-weight.
- Terminal-Bench 3.0 (agentic terminal/coding tasks): jumped from 4.6 to 28.3, a 6.2x improvement over its predecessor — the single largest reported jump among this batch of results.
- DeepSWE 1.1 (software engineering benchmark): 66.9%.
- CyberGym (cybersecurity capability benchmark): 84.5%, narrowly ahead of Anthropic's Mythos 5 (83.8%) and OpenAI's GPT-5.6 Sol (83.6%) — notable because it's an open-weight model outscoring two closed frontier labs on this specific benchmark.
- The scale of the Terminal-Bench jump (6x) relative to more modest gains on DeepSWE and CyberGym suggests targeted post-training on agentic terminal/coding tasks specifically, rather than a uniform across-the-board capability increase.
Reel Script
Hook: An open-weight Chinese model just beat both Anthropic and OpenAI's newest models on a cybersecurity benchmark — and it's free to download.
Core Concept: GLM-5.3 is a mixture-of-experts model — meaning instead of one giant network processing every query, it's built from many smaller specialist sub-networks, and a router picks which ones to activate per request, so you get the capacity of a huge model without paying the full compute cost on every token. What makes this release notable isn't the architecture, though — it's open-weight, meaning anyone can download and run it themselves, competing directly against closed models you can only access through an API.
Hands-On: Look at the actual benchmark movement. On Terminal-Bench 3.0, which tests an agent's ability to complete real tasks inside a terminal environment, GLM-5.3 scored 28.3 versus its predecessor's 4.6 — a 6.2x jump, the kind of gain that usually only comes from targeted post-training rather than general scaling. On DeepSWE 1.1, a software-engineering benchmark, it hit 66.9%. But the number that should get attention is CyberGym, a benchmark for cybersecurity task capability: GLM-5.3 scored 84.5%, ahead of Anthropic's Mythos 5 at 83.8% and OpenAI's GPT-5.6 Sol at 83.6%. The gap is narrow — under a point over both — but the fact that an open-weight model is even in that conversation, let alone narrowly ahead, is the real signal.
Takeaway: Open-weight models closing the gap on frontier closed models isn't new, but leading on a security-relevant benchmark changes the calculus for anyone deciding whether to self-host versus pay for API access — especially for cybersecurity or coding-agent use cases where GLM-5.3's specific gains land hardest. Worth benchmarking against your own workload before assuming the closed model is still the safer default.