GLM-5: The 754-Parameter Giant and the Rise of Agentic Engineering

China’s largest open-weight AI model raises the stakes — and coins a new term for AI-era software development

If you thought the AI model size wars were cooling down, think again. Z.ai just released GLM-5, a 754-billion parameter model with 1.51 terabytes of weights on Hugging Face — roughly twice the size of GLM-4.7.

But the model itself isn’t the most interesting part. The announcement also marks a pivotal moment in how we talk about AI-assisted software development.

The Core Insight

GLM-5 is MIT-licensed and available on Hugging Face, making it one of the largest open-weight models ever released. Simon Willison tested it with his signature “pelican on a bicycle” SVG prompt and got “a very good pelican on a disappointing bicycle frame” — a characteristically honest assessment.

But beyond the benchmarks, the announcement is notable for its framing: “From Vibe Coding to Agentic Engineering.”

This terminology isn’t accidental. The term “agentic engineering” has been gaining traction among AI industry leaders, notably:

• Andrej Karpathy
• Addy Osmani

It’s a deliberate pivot from “vibe coding” — where developers describe what they want and let AI figure it out — to a more intentional, engineering-focused discipline.

Why This Matters

The shift from “vibe coding” to “agentic engineering” reflects a maturing understanding of how AI fits into software development:

Vibe Coding (current state):
– Describe the outcome you want
– Iterate on the results
– AI as a creative partner
– Fast but potentially uncontrolled

Agentic Engineering (emerging):
– AI as an autonomous agent with clear objectives
– Structured workflows and verification
– Human oversight at boundaries
– More predictable and maintainable

The GLM-5 release represents both a capability milestone (largest open Chinese AI model) and a conceptual milestone (framing AI’s role in development as engineering rather than magic).

Key Takeaways

• 754B parameters — More than double GLM-4.7’s 368B
• MIT licensed — Open weights for commercial use
• 1.51TB total size — Requires serious hardware to run
• “Agentic Engineering” emerging — Industry coalescing on terminology
• Open-source Chinese AI — Growing competitive pressure on Western labs

Looking Ahead

The term “agentic engineering” signals a maturing of expectations. As AI capabilities grow, the challenge shifts from “what can AI do?” to “how do we build reliable systems with AI?”

This framing matters because it acknowledges both the power and the limitations: AI can do remarkable things, but integrating it into software engineering requires discipline, testing, and architectural thinking.

GLM-5 may or may not become a dominant model. But the vocabulary shift it represents — from vibes to engineering — is likely here to stay.

Based on analysis of Z.ai’s GLM-5 announcement and Simon Willison’s coverage