GLM-5

GLM-5 is a large open-weights mixture-of-experts model positioned for complex systems engineering and long-horizon agentic work rather than casual code generation. The architecture scales from the prior generation's 355B parameters with 32B active to 744B parameters with 40B active, supported by a pre-training corpus expanded from 23T to 28.5T tokens. To keep deployment economical at this scale, GLM-5 integrates DeepSeek Sparse Attention, which compresses long-context computation without sacrificing the model's ability to handle extended inputs. Together, these changes mark a deliberate step from snippet-level code assistance toward end-to-end engineering, where the model is expected to plan, coordinate tools, and sustain multi-step tasks across long contexts.

On the post-training side, the team paired a new asynchronous reinforcement-learning infrastructure called slime with continued scale, enabling more granular iterations on long-horizon task environments. The result is broad gains on academic benchmarks relative to earlier GLM variants, with the model also surfaced through developer-friendly channels such as Novita AI's Model APIs and Amazon Bedrock, confirming production availability for teams that want to wire it into agentic pipelines. Practical fit is strongest for organizations building software agents, automation backends, or research prototypes that need a capable, open-weights base they can self-host or route through managed inference, and who value long-context reasoning paired with sparse-attention efficiency over a smaller, denser alternative.