Huawei Open-Sources openPangu-2.0-Pro, First 500B+ Model Trained on Non-NVIDIA Hardware

Huawei open-sourced openPangu-2.0-Pro, a 505B-parameter MoE model with 18B active parameters and 512K context window, trained entirely on Huawei Ascend NPUs. Benchmark comparisons show it in the 'second tier' among Chinese open-source models: GPQA-Diamond 87.9 (vs Kimi K3 93.5, GLM-5.2 91.2), BrowseComp 65.7 (vs K3 91.2), SWE-bench Verified 68.5 (similar to DeepSeek V3.1). The author notes that Huawei's evaluation dimensions barely overlap with DeepSeek's, making direct comparison difficult, and that the model's parameter count is several times smaller than Kimi K3 and GLM-5.2, so the performance is expected.