Zhipu (Z.ai) releases GLM-5, trained entirely on Huawei Ascend chips

Zhipu, trading publicly as Z.ai after a January 2026 Hong Kong IPO, released GLM-5, a 744-billion-parameter mixture-of-experts model with 40 billion active parameters per token, roughly double the scale of its predecessor GLM-4.5. The company said the model was trained on 28.5 trillion tokens, adopted DeepSeek’s sparse-attention mechanism for a 200,000-token context window, and reported scores including 77.8% on SWE-bench Verified and 86.0% on GPQA-Diamond. Weights were released under the permissive MIT licence.

The training-hardware claim drew more attention than the benchmarks. Reporting attributed to Reuters, and cited in independent analysis of the release, said GLM-5 was trained entirely on Huawei Ascend chips using Huawei’s MindSpore framework, with no dependence on Nvidia hardware — a claim some coverage said Z.ai had not itself directly confirmed. Separate reporting put the training cluster at around 100,000 Ascend 910B processors, chips designed by Huawei’s HiSilicon unit and fabricated domestically by SMIC. If accurate, it would be the first frontier-scale model trained without any US-controlled chips, a marker of progress for China’s effort to build an AI compute stack independent of export-controlled Nvidia hardware — relevant given Zhipu’s own placement on the US Commerce Department’s Entity List since January 2025.

GLM-5 arrived within days of a cluster of other major Chinese model releases timed ahead of the Lunar New Year, including Alibaba’s Qwen 3.5 and ByteDance’s Doubao 2.0 and SeeDance 2.0, with DeepSeek’s next model expected in the same window. Coverage treated the release less as a single capability jump than as evidence that Chinese labs could sustain frontier-model pace on domestic hardware even as US export controls tightened, though independent verification of some of Zhipu’s benchmark claims remained outstanding at the time of release.