Zhipu AI Releases GLM-5 Open-Source Model Trained Entirely on Huawei Chips, Shares Surge 34% as Benchmarks Rival Claude Opus 4.5

Zhipu AI released GLM-5 on February 11, 2026, a 744 billion parameter large language model trained entirely on domestically produced Huawei Ascend chips without any Nvidia hardware, triggering a 34 percent stock surge on the Hong Kong Stock Exchange as the open-source model achieved benchmark performance rivaling Anthropic's Claude Opus 4.5 at a fraction of the cost.

The launch represents what industry analysts describe as a generational leap in China's AI capabilities and a milestone in the country's drive toward self-reliant AI infrastructure amid ongoing US semiconductor export restrictions. The model's competitive performance on coding and reasoning benchmarks positions Chinese AI companies as genuine innovators rather than fast followers for the first time.

GLM-5 utilizes a Mixture-of-Experts architecture with 744 billion total parameters and 44 billion active parameters per inference, representing more than double the scale of predecessor GLM-4.5's 355 billion parameters. The model was trained on 28.5 trillion tokens using Zhipu's novel "slime" asynchronous reinforcement learning infrastructure that substantially improves training throughput and efficiency compared to traditional approaches.

Performance benchmarks demonstrate parity with leading proprietary Western models across multiple dimensions. On SWE-bench Verified measuring real-world coding capabilities, GLM-5 achieved 77.8 percent, ranking first among all open-source models though trailing Claude Opus 4.5's 80.9 percent and GPT-5.2's 80.0 percent. On Humanity's Last Exam with tools enabled, GLM-5 scored 50.4, outperforming Claude Opus 4.5's 43.4 and approaching GPT-5.2's 45.5.

VentureBeat reported GLM-5 achieved a record-low hallucination rate on the independent Artificial Analysis Intelligence Index with a score of negative one, representing a 35-point improvement over its predecessor and leading the entire AI industry including Google, OpenAI, and Anthropic in knowledge reliability by knowing when to abstain rather than fabricate information.

The model's training exclusively on Huawei Ascend chips using the MindSpore framework demonstrates China's semiconductor ecosystem can now support compute-intensive frontier AI development at scale despite US export restrictions on advanced chips. Zhipu also optimized GLM-5 for deployment on other domestic Chinese processors including Moore Threads, Cambricon, Kunlun Chip, MetaX, Enflame, and Hygon through kernel optimization and model quantization.

Zhipu AI released GLM-5 under the permissive MIT license enabling unrestricted commercial use, adaptation, and redistribution, directly challenging Western labs' tendency toward closed or restrictively licensed frontier models. Model weights are publicly accessible on Hugging Face and ModelScope with support for inference frameworks including vLLM and SGLang for local implementations.

Pricing represents aggressive market positioning at approximately 0.80 dollars per million input tokens and 2.56 dollars per million output tokens on OpenRouter as of February 11, roughly six times cheaper on input and nearly ten times cheaper on output than Claude Opus 4.6's five dollars input and 25 dollars output pricing.

The 34 percent stock surge following launch reflects growing investor confidence in Chinese AI companies' ability to challenge entrenched Western competitors. The rally extended beyond Zhipu, lifting the broader index of Chinese AI stocks as market analysts validated the "China Speed" narrative positioning domestic firms as genuine innovators capable of setting pace rather than following.

GLM-5 features native Agent Mode capabilities transforming raw prompts directly into professional office documents including formatted Word files, PDFs, and Excel spreadsheets with proper styling delivered as ready-to-use outputs. The model integrates DeepSeek Sparse Attention preserving 200,000 token context capacity while drastically reducing deployment costs.

Prior to official announcement, GLM-5 appeared under the alias "Pony Alpha" on OpenRouter in early February processing 40 billion tokens on day one, allowing developers to evaluate performance without brand bias before the formal reveal validated competitive capabilities against Western frontier models.