Huawei is ramping up its independent artificial intelligence (AI) ecosystem by unveiling the source – the core technology – of a colossal AI model with over 500 billion parameters. The tech giant seems to be constructing its AI model on its proprietary neural processing unit (NPU), reducing external reliance, and aims to swiftly boost its clout in the global AI market through open-source strategies.
On Tuesday, Huawei Central reported that Huawei recently released the core components of its large language model (LLM), openPangu 2.0 Pro, as open-source. This includes model weights, basic inference code, and a technical report.
The weights represent the AI’s core intelligence, developed at a cost of hundreds of billions of KRW (hundreds of millions of USD) and months of infrastructure investment. With the weights now public, developers and companies can download them directly to their servers, allowing them to build systems without the massive upfront costs typically associated with training large-scale AI.
The provided basic inference code allows developers to immediately run the model and integrate it into their services or internal systems. The technical report serves as a foundation for other researchers to understand the model’s mechanics and develop improved versions.
The newly released openPangu 2.0 Pro boasts a staggering 505 billion parameters. In terms of this AI knowledge repository, it outpaces Meta’s latest open-source model, Llama 3.1, which has 405 billion parameters.
The model supports a context length of 512K, meaning it can process up to 512,000 tokens at once. This translates to about 350,000 to 400,000 characters in Korean or roughly 380,000 to 400,000 words in English.
However, it falls short when compared to Google’s latest flagship AI, Gemini 3.5 Pro, which handles 2 million tokens, operating at about 25% of that capacity. It’s also at about 50% capacity compared to OpenAI’s ChatGPT 5.5, which manages around 1.05 million tokens. This places Huawei’s model at a smaller scale compared to global leaders in the ultra-large context race.
Huawei’s decision to open-source its code likely stems from a strategic move towards semiconductor independence and ecosystem dominance.
Unlike many companies relying on NVIDIA’s graphic processing unit (GPU) infrastructure for AI operations, Huawei completed the entire process – from pre-training to inference and post-training – using its own NPU chipset, Ascend.
Amid U.S. technology sanctions limiting access to cutting-edge semiconductors, Huawei is swiftly locking in global and local corporate customers and developers. It’s leveraging its comprehensive high-tech component ecosystem and tailored AI platform to achieve this, according to industry analysts.
Huawei Central reports that the company has urged developers, corporate partners, and researchers worldwide to actively use this model. They’re also soliciting diverse feedback and suggestions to ensure the continued growth of the open-source ecosystem.