Paper summary: GLM-5: from Vibe Coding to Agentic Engineering

Problem

Following GLM-4.5 (summarised here in Aug 2025), the goal moves from “vibe coding”, meaning casual AI-assisted coding, to agentic engineering: models that carry out long, end-to-end software tasks.

Method

• DSA (DeepSeek Sparse Attention, used during continued pre-training per the paper) reduces training and inference cost while keeping long-context fidelity.
• A new asynchronous RL infrastructure decouples generation from training, for much more efficient post-training.
• New asynchronous agent RL algorithms help the model learn from complex, long-horizon interactions.

Key results

From the abstract: state-of-the-art performance on major open benchmarks, and “unprecedented” capability on real-world, end-to-end software engineering. Figure 1 compares GLM-5 with DeepSeek-V3.2, Claude Opus 4.5, Gemini 3 Pro and GPT-5.2 on eight agentic, reasoning and coding benchmarks, including Humanity’s Last Exam and SWE-bench Verified.

Limitations

There’s no dedicated limitations section in the extracted text. Comparisons are the developers’ own.