Huawei open-sources OpenPangu-2.0-Flash with 512K context
92B total, 6B active MoE model now available with full code.
Huawei has officially open-sourced OpenPangu-2.0-Flash, a large language model from their OpenPangu 2.0 series. The Flash variant uses a Mixture-of-Experts (MoE) architecture with 92 billion total parameters, but only 6 billion are active per forward pass—making it efficient for deployment. It supports a massive 512,000-token context window, enabling handling of long documents, codebases, or multi-turn conversations.
Alongside weights, Huawei released the inference code and training operations, allowing developers to fine-tune and run the model on their own infrastructure. A larger Pro variant, with 505 billion total parameters (18B active), is scheduled for release in July. Huawei plans to release more open-source components later this year, signaling a strong push to compete with other open-weight models like Llama and Qwen.
- OpenPangu-2.0-Flash: 92B total, 6B active MoE model open-sourced by Huawei
- Supports 512K token context length, among the longest for open models
- Pro variant (505B total, 18B active) arriving July 2025; more components later this year
Why It Matters
Huawei's open-source release lowers entry barriers for long-context AI applications, challenging Western models in efficiency and scale.