GLM-5: the model family and the model that started it

GLM-5 was listed on OpenRouter on 11 February 2026. Z.ai describes it as a model "targeting complex systems engineering and long-horizon agentic tasks" — in other words, it is meant to work through big, multi-step programming jobs rather than to be a quick chat companion.

Its model card gives the architecture in plain numbers. Compared with GLM-4.5, GLM-5 scales from 355 billion parameters (32 billion active) to 744 billion parameters with 40 billion active per token, and its pre-training data grows from 23 trillion to 28.5 trillion tokens. Only a fraction of the weights is used for each token, which is what keeps a model this large practical to serve.

GLM-5 also adopts DeepSeek Sparse Attention (DSA). Sparse attention lets the model look at a selected subset of earlier tokens instead of all of them, which Z.ai says largely reduces deployment cost while keeping long-context ability. On OpenRouter the original GLM-5 has a context window of about 205K tokens and can return up to about 131K tokens in one answer.

For post-training, Z.ai built slime, an asynchronous reinforcement-learning infrastructure that it says improves training throughput and allows more fine-grained iterations. The model card reports results across reasoning, coding and agent benchmarks — for example 77.8 on SWE-bench Verified and 56.2 on Terminal-Bench 2.0 with the Terminus 2 harness — and claims best-in-class performance among open-source models at release.

Those are Z.ai's own numbers measured with its own harnesses. They are useful for comparing GLM versions with each other, but the honest test is your own task: paste a real problem into the chat and judge the answer.

After GLM-5, Z.ai shipped a steady line of updates. GLM-5 Turbo (March 2026) and the multimodal GLM-5V-Turbo (April 2026) target fast agent workflows through the API. GLM-5.1 (April 2026) focused on long-horizon coding. GLM-5.2 (June 2026) brought a solid one-million-token context and adjustable reasoning effort. GLM-5.3 (August 2026) kept the GLM-5.2 base and improved post-training, and GLM-5.3 Flash added native image and video input at a much lower price. GLM-5.3 FlashX and GLM-5.3 Prime are high-speed API variants.

Five of these models — GLM-5, GLM-5.1, GLM-5.2, GLM-5.3 and GLM-5.3 Flash — are published as open weights on Hugging Face. The datasheet below and the family table further down list every model with its context window, price and licence.

For new projects, usually not. Later models in the family have a larger context window, adjustable reasoning and, in the case of GLM-5.3 Flash, far lower prices. The original GLM-5 is still interesting if you want a fully MIT-licensed open model with a published architecture and paper (arXiv 2602.15763), or if you need to reproduce results that were measured on it.