Zhipu AI, China's first publicly traded AI company, released GLM-5 in February 2026, a 744-billion-parameter open-source model trained entirely on Huawei Ascend chips. The model demonstrates that frontier AI systems can achieve GPT-5.2-level performance without NVIDIA hardware while addressing one of the field's most persistent challenges: hallucinations.
GLM-5 Reduces Hallucinations From 90% to 34%
The most significant technical breakthrough in GLM-5 is its Slime RL (Reinforcement Learning) technique, which reduced hallucinations from 90% in the previous GLM-4.7 model to 34% in GLM-5. This 62% improvement represents one of the largest documented hallucination reductions between successive model generations. The Slime RL approach outperforms traditional methods for improving model reliability, addressing a critical barrier to deploying large language models in production environments where factual accuracy is essential.
Training on 100,000 Huawei Ascend Chips
GLM-5 was trained using 100,000 Huawei Ascend 910B chips with the MindSpore framework, making it the first frontier model of this scale built entirely outside the NVIDIA ecosystem. This achievement demonstrates China's growing capability to develop cutting-edge AI models using domestic hardware, reducing dependence on Western semiconductor technology. According to multiple reviews, GLM-5 rivals GPT-5.2 performance on numerous benchmarks despite being trained on zero NVIDIA silicon.
Cost Efficiency and Open-Source Availability
GLM-4.7, the predecessor model, costs $0.11 per million input tokens compared to Claude Opus at $15, offering a 136x cost advantage for production deployments. GLM-5 continues this cost-effective approach while being fully open-weight, with model weights available on Hugging Face and API access through multiple providers. The open-source release enables researchers and companies worldwide to build upon the model without licensing restrictions.
Market Impact and Evolution
Zhipu AI's GLM repository doubled its forks and merged pull requests within three months in 2026, indicating strong developer adoption. The latest version, GLM-5.1, is designed specifically for agentic engineering and complex, long-horizon software development tasks. The model's success contributes to a broader trend where open-source models like Mistral 3, Llama 4, and DeepSeek's offerings are now competitive with proprietary systems for a wide range of business applications.
Key Takeaways
- GLM-5 reduced hallucinations from 90% to 34% using a novel Slime RL technique, representing a 62% improvement over GLM-4.7
- The 744-billion-parameter model was trained entirely on 100,000 Huawei Ascend 910B chips without any NVIDIA hardware
- GLM-4.7 costs $0.11 per million input tokens compared to Claude Opus at $15, offering 136x cost savings
- Model weights are fully open-source and available on Hugging Face, with API access through multiple providers
- GLM-5 achieves performance comparable to GPT-5.2 on multiple benchmarks, demonstrating that open-source quality is converging with proprietary models