openPangu-2.0-Pro

model
Huawei

Huawei's open-weight flagship: a ~505B-total / ~18B-active MoE with 512K context, trained on ~34T tokens — claimed as the first ~500B-class model fully pretrained on Ascend 910B with no NVIDIA hardware anywhere in the run. Released under the OPENPANGU MODEL LICENSE 2.0 via GitCode and ModelArts (not HuggingFace). Continues the compute-sovereignty arc alongside GLM-5 (Ascend 910B) and DeepSeek-V4 (Ascend 950PR).

Active-parameter figure per the primary GitCode page (~18B); some press reports "180B active" — treat as an error.

GitCodeCoverage (Pandaily)

Model Details

Architecture MOE
Parameters 505B
Active params 18B
Context window 512,000
Training tokens 34T
License OPENPANGU MODEL LICENSE 2.0
frontiermoeopen-weight

Related