GLM-5

What it is

GLM-5 is a large language model from Zhipu (智谱), released on 12 February 2026 under the MIT License.^[2][1] Its weights were published as an open model, and its technical report carries the title "GLM-5: from Vibe Coding to Agentic Engineering" (arXiv 2602.15763).^[1]

Origins

GLM-5 scales up from GLM-4.5, which has 355 billion total parameters and 32 billion active. GLM-5 has 744 billion total parameters (744B) and 40 billion active.^[1]

Main features

Pre-training data grew from 23 trillion to 28.5 trillion tokens. Pre-training is the stage in which a model learns from a large body of text before any task-specific tuning.^[1]

GLM-5 integrates DeepSeek Sparse Attention (DSA), which the model uses to cut deployment cost while keeping its long-context capacity. Sparse attention lets a model focus on a selected subset of earlier tokens rather than all of them, which lowers the computation needed for long inputs.^[1]

Post-training used slime, Zhipu's asynchronous reinforcement-learning infrastructure. Reinforcement learning (RL) trains a model by rewarding good results, and asynchronous training lets the parts of that process run without waiting for one another.^[1]

The model card includes a deployment guide for Ascend NPUs, neural processing units.^[1]

On SWE-bench Verified GLM-5 scored 77.8, and on SWE-bench Multilingual it scored 73.3. On Terminal-Bench 2.0 it scored 56.2 and 60.7. On HLE it scored 30.5, and 50.4 with tools. On τ²-Bench it scored 89.7, and on BrowseComp it scored 75.9 with context management.^[1]

Zhipu claimed that GLM-5 had the best performance among open-source models on reasoning, coding and agentic tasks.^[1]

Its place in development

In February 2026 Zhipu's shares fell about 23% amid compute shortages and user complaints, and new sign-ups were restricted.^[2] GLM-5 was succeeded by GLM-5.1, which arrived between March and April 2026.^[2]

Reading the benchmark figures

A benchmark score is only meaningful when two models are run on the same test under the same conditions. Vendors usually publish comparison tables that set their model against selected rivals on selected tests, so a number should be read together with the name of the test and the models it is compared with. A small gap between two scores does not by itself show a real difference in ability, and one test rarely captures everything a model can do.