openPangu 7B and Pangu Pro MoE (72B)
Huawei open-sourced a 7B model and the 72B Pangu Pro MoE (16B active), built for Ascend; a GitHub study then alleged overlap with Qwen.
MoGE forces equal expert activation per device group to balance load; the paper reports 1,148 tokens/s per Ascend card (1,528 with speculative decoding). On 2025-07-04 researchers alleged high similarity to Alibaba's Qwen; Huawei's Noah's Ark Lab denied incremental training on another model.
- Date
- Monday, 30 June 2025
- Lab
- Huawei (Pangu)
- Kind
- open-weights
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total / active parameters | 72B / 16B | company |
| Inference throughput | 1,148 tokens/s per card Ascend; 1,528 with speculative decoding | company |
Open-sourcing on 2025-06-30 and the 2025-07-04/05 dispute are from Wikipedia; the paper itself is arXiv 2025-05-27. Whether similarity reflects copying is disputed; I did not read the fingerprint study or Huawei's statement.
Sources
This record was checked against its sources on 6 October 2026. How we check