Z.ai (formerly Zhipu AI) released GLM-5.1 on March 27, 2026, a 744-billion-parameter open-source model trained entirely on Huawei Ascend 910B chips. The release gained viral attention on April 7 when performance comparisons showed the model achieving 94.6% of Claude Opus 4.6's coding capabilities at $3 per month, demonstrating China's progress in building frontier AI capabilities independent of US semiconductor exports.
GLM-5.1 Uses Mixture-of-Experts Architecture With 256 Experts
GLM-5.1 employs a Mixture-of-Experts (MoE) architecture with 744 billion total parameters across 256 experts, activating 8 experts per token for 40 billion active parameters during inference. The model was trained on a cluster of 100,000 Huawei Ascend 910B accelerator chips, representing one of the largest non-NVIDIA training runs globally. Z.ai released the model under an MIT license on Hugging Face, making the weights freely available for commercial use.
Model Achieves Near-Frontier Performance on Coding Benchmarks
According to Z.ai's published benchmarks, GLM-5.1 scored 52.3% on HLE (with tools), surpassing GPT-5.4's 52.1% and approaching Claude Opus 4.6's 54.9%. On specialized agentic coding tasks, the model achieved 45.3 on an internal scoring metric compared to Claude Opus 4.6's 47.9, representing 94.6% of the frontier model's performance. Z.ai also claims state-of-the-art performance among open-source models on SWE-Bench Pro and Terminal-Bench, though independent verification of these benchmarks is pending.
Training on Huawei Chips Demonstrates Sovereign AI Capabilities
Zhipu AI was added to the US Entity List in January 2025, restricting the company's access to NVIDIA GPUs and other advanced US semiconductor technology. The successful training of GLM-5.1 on Huawei's domestic silicon demonstrates significant progress in China's efforts to build competitive AI infrastructure without US components. This development has major geopolitical implications for the global AI industry, showing that export controls alone may not prevent the development of frontier AI capabilities.
API Access Available at $3 Monthly Through Z.ai Coding Plan
Z.ai offers API access to GLM-5.1 through a "Coding Plan" priced at $3 per month, significantly undercutting Western frontier models. The pricing strategy has generated discussion about competitive dynamics in the global AI market, particularly for coding applications where GLM-5.1 appears to offer near-frontier performance at a fraction of typical costs.
Key Takeaways
- GLM-5.1 is a 744-billion-parameter MoE model with 256 experts and 40 billion active parameters per inference, released under MIT license
- The model was trained entirely on 100,000 Huawei Ascend 910B chips, demonstrating China's ability to build frontier AI without US semiconductor access
- GLM-5.1 scored 52.3% on HLE benchmarks, beating GPT-5.4 and achieving 94.6% of Claude Opus 4.6's coding performance according to Z.ai
- API access costs $3 per month through Z.ai's Coding Plan, significantly undercutting Western frontier model pricing
- Zhipu AI has been on the US Entity List since January 2025, restricting access to NVIDIA GPUs and making this release a demonstration of sovereign AI capabilities