GLM-5.2 (huggingface.co)

The latest release from GLM, the GLM-5.2 model, represents a significant advancement for long-horizon tasks, boasting a stable 1 million token context for the first time. This upgrade enhances the model's capability for prolonged tasks, aligning better with real-world applications where extended input is crucial. Significant technical improvements include the introduction of IndexShare to reduce per-token FLOPs by 2.9× across sparse attention layers and enhancements to the speculative decoding mechanism in the MTP layer, which increases the acceptance length by up to 20%. The model’s advanced coding capabilities allow for varying levels of computational effort, optimizing performance while balancing latency, making it an attractive option for developers and researchers focused on coding-related tasks. GLM-5.2 is released under an MIT open-source license, meaning researchers globally can access and build upon this state-of-the-art model without regional restrictions. With benchmark scores showing marked improvements over its predecessor and other competitive models, GLM-5.2 is poised to play a crucial role in the evolution of AI tools for coding and long-horizon reasoning.