GLM-5 Shatters Records: First Frontier Model on Huawei Chips Tops HLE at 50.4%!
The 744B-parameter model runs on Huawei Ascend chips, costs 136x less than Claude Opus 4.5, and reduces hallucinations to 1.2%.
Zhipu AI's GLM-5 has emerged as a landmark achievement in the AI industry, becoming the first frontier-class model trained entirely on Huawei Ascend chips without using a single NVIDIA GPU. The 744-billion-parameter model achieved the top score on the HLE benchmark at 50.4%, demonstrating that competitive AI development is possible outside the traditional NVIDIA-dominated hardware ecosystem. Released under an MIT license at just $0.11 per million tokens—approximately 136x cheaper than Claude Opus 4.5—GLM-5 represents both a technical and strategic breakthrough in the global AI race.
The model's technical innovations include the Slime RL technique, which reduces hallucination rates to just 1.2%, making it three times more reliable than competing models. This combination of benchmark performance, cost efficiency, and reduced hallucinations positions GLM-5 as a viable alternative to proprietary models from OpenAI, Anthropic, and Google. The successful deployment on Huawei hardware signals a major shift in AI infrastructure independence, potentially reshaping global supply chains and reducing reliance on Western chip technology for advanced AI development.
- First frontier model trained entirely on Huawei Ascend chips (no NVIDIA GPUs)
- Achieves #1 HLE benchmark score of 50.4% with 744 billion parameters
- Reduces hallucination rates to 1.2% using Slime RL technique and costs $0.11 per million tokens
Why It Matters
Breaks NVIDIA's GPU monopoly for frontier AI, offering cost-effective, reliable models that reduce Western technology dependencies.