GLM-5: from Vibe Coding to Agentic Engineering

A shift in what coding models are for

The GLM-5 technical report is credited to the GLM-5-Team at Zhipu, with Aohan Zeng, Xin Lv and Zhenyu Hou leading the author list, and was submitted to arXiv on 17 February 2026 as 2602.15763 ^[1]. It frames GLM-5 as a generation of foundation model meant to move the paradigm from "vibe coding" to "agentic engineering" ^[1]. The difference is about what the model is asked to do. Vibe coding means taking generated code largely as it comes, one prompt at a time; agentic engineering means the model carries a longer engineering task through — planning, editing, running, checking — more like an agent than a single-response assistant. The report builds on the agentic, reasoning and coding (ARC) capabilities of its predecessor ^[1].

Sparse attention and a large mixture of experts

The central architectural choice is DSA, DeepSeek Sparse Attention, which allocates attention by token importance ^[1]. Attention is normally computed against every earlier token, so cost grows with context length; allocating it according to importance sharply cuts training and inference cost while keeping long-context fidelity ^[1]. With DSA the parameter count scales to 744 billion, with 40 billion active per token, in a mixture-of-experts architecture ^[1]. In a mixture-of-experts model, a router sends each token to a small subset of the network rather than the whole of it, so the total number of parameters a model holds and the amount of computation spent on any one token can grow separately. That is how a model of this size can run at a per-token cost set by the 40 billion active parameters.

Post-training with asynchrony

The report introduces a new asynchronous reinforcement learning infrastructure that decouples generation from training, in order to raise post-training efficiency ^[1]. It also proposes asynchronous agent RL algorithms, so that the model can learn from complex long-horizon interactions ^[1]. The motivation is structural: training an agent by reinforcement learning is slow because the model has to act through many steps before any reward arrives, and if the trainer sits idle while the model generates, most of the wall-clock time is wasted. Separating generation from training lets the two sides of the loop run at once.

Results

GLM-5 is reported to reach state of the art on major open benchmarks ^[1]. The report highlights real-world coding in particular, where GLM-5 surpasses previous baselines on end-to-end software engineering tasks ^[1]. Code and models are published with the report ^[1].

Why it matters

The interest of the work is less in a single score than in the pair of claims the report puts side by side. One is about capability: the target is no longer autocompleting a function but finishing an engineering task end to end, which is what the benchmark choice and the agent RL algorithms are aimed at. The other is about cost: sparse attention and a mixture-of-experts layout are the levers used to keep a 744-billion-parameter model trainable and servable, with 40 billion parameters active per token.

Publishing code and models alongside the report means the architecture and training setup can be inspected and reused rather than taken on trust ^[1]. For readers trying to follow where coding models are heading, the report is a compact statement of one direction: longer autonomous work, paid for by cheaper attention and by post-training pipelines that do not wait on themselves.