GLM 5

GLM 5 is the February 2026 flagship from Z.ai (formerly Zhipu AI), positioned for complex systems engineering and long-horizon agentic work. Architecturally it is a Mixture-of-Experts model with 744 billion total parameters and roughly 40 billion active per token, scaled up from GLM-4.5's 355B (32B active), with pre-training data expanded to 28.5 trillion tokens. It is distributed open-weight under the MIT license in both full-precision and FP8 formats.

The two headline changes over earlier generations are efficiency-focused. GLM 5 adopts DeepSeek Sparse Attention (DSA), which the technical report describes as dynamically allocating attention by token importance to lower compute without compromising long-context understanding — an advance over the standard MoE used in GLM-4.5. Post-training uses a new asynchronous reinforcement-learning infrastructure built on the "slime" framework that decouples generation from training to improve GPU utilization.

Relative to its same-family predecessor GLM 4.7, Z.ai reports significant improvements across academic benchmarks in reasoning, coding, and agentic tasks.

GLM 5 was followed by refreshed siblings GLM 5.1 and GLM 5.2, the latter extending to a roughly 1M-token context with the IndexShare architecture.

Sources

GLM-5 - Overview - Z.AI DEVELOPER DOCUMENT· docs.z.aiGLM-5: from Vibe Coding to Agentic Engineering· arxiv.orgzai-org/GLM-5 · Hugging Face· huggingface.co

This About section is AI-generated from public sources via VeniceStats + Venice inference, with no human editing. It may contain inaccuracies.