Zhipu AI releases open-source GLM-5.1 model with coding performance rivaling proprietary leaders

Zhipu AI has released GLM-5.1, an open-source large language model that achieves performance on coding tasks competitive with leading proprietary models, marking a significant milestone for open-weights AI development.^[1][5] The model, released under a permissive MIT License, scores 77.8% on SWE-bench Verified—only 3 points behind Claude Opus 4.6 (80.8%) and GPT-5.2 (80.0%)—while claiming 94.6% of Claude Opus 4.6's overall coding performance.^[5]

GLM-5.1 represents a focused refinement of GLM-5 rather than a complete architectural overhaul.^[5] The post-training optimization concentrated specifically on coding capabilities, achieving a 28% improvement in coding benchmark scores, which jumped from 35.4 (GLM-5) to 45.3 (GLM-5.1).^[5] This enhancement was accomplished through progressive alignment techniques including multi-task supervised fine-tuning, reasoning reinforcement learning, agentic reinforcement learning, and on-policy cross-stage distillation.^[5]

The model retains GLM-5's substantial scale, featuring approximately 744 billion total parameters with 40 billion active during inference, and maintains a 200,000-token context window.^[6] These specifications position it for extended, complex tasks—what developers describe as "long-horizon" work.^[6][7] According to early assessments, GLM-5.1 shows marked improvement in instruction following, debugging, and maintaining focus on primary objectives compared to its predecessor, though it has become noticeably more code-centric, sometimes using code or HTML even when simpler text answers would suffice.^[7]

The open-source release carries significant implications for enterprise adoption and cost efficiency.^[1][5] Organizations can now download, customize, and deploy the model for commercial purposes without licensing restrictions.^[1] Pricing analyses suggest GLM-5.1 operates at approximately $1.00 to $3.20 per million tokens, positioning it as a cost-competitive alternative to proprietary frontier models.^[5] However, practical deployment considerations remain: the model requires over 1 terabyte of local memory to run, which may result in substantial hardware costs despite lower per-token pricing compared to API-based alternatives.^[1]

The release underscores growing competition in the open-source AI landscape, particularly from Chinese AI developers.^[1] While GLM-5.1 is described as a frontier-adjacent model offering compelling value for coding and agentic tasks, important limitations persist—notably its text-only nature, lacking multimodal capabilities such as image or video input that some competing open models provide.^[1] The weights for GLM-5.1 have been promised but were not yet publicly released at the time of announcement, though GLM-5's weights remain available on HuggingFace under the MIT license.^[5]