Huawei Opensources 505 Billion-Parameter openPangu-2.0-Pro Model, Weights and Inference Code Are Released Simultaneously

On July 31, Huawei officially open-sourced the large model openPangu-2.0-Pro under the openPangu AI model brand openPangu, along with model weights, basic inference code, and technical reports, further promoting the construction of the Ascend AI ecosystem and providing best practices for developers and enterprises based on native Ascend training and inference technologies.

According to the information, openPangu-2.0-Pro is a large-scale Mixture of Experts (MoE) language model trained on Ascend NPU, with a total parameter scale of approximately 505B (505 billion), about 18B parameters activated per token, supporting a context length of 512K, and a training data size of about 34T Tokens. In the post-training phase, the model completed supervised fine-tuning (SFT) with fast and slow integration, multiple specialized reinforcement learning (RL), and achieved capability fusion through online distillation (OPD), further enhancing its overall performance.

This open-source release is an important part of Huawei's openPangu 2.0 open-source plan. At the HDC 2026 Developer Conference held in June, Huawei Executive Director, Chairman of the Product Investment Review Committee, and President of the Terminal BG, Yu Chengdong, announced that openPangu 2.0 would gradually open-source seven key components starting from June 30, including pre-training code, post-training code, and training operators. Previously, the 92B-parameter openPangu-2.0-Flash model was released first on June 30.

Regarding the total parameter scale of openPangu-2.0-Pro being 505B, Yu Chengdong stated that Huawei will allocate a large amount of AI computing resources to support the domestic industry ecosystem. However, Huawei's available computing power is relatively limited, and AI training costs continue to be high. Therefore, Huawei pays more attention to optimizing the model's efficiency in terms of latency and throughput. As openPangu 2.0 continues to open up core models and training capabilities, the Ascend-native AI development ecosystem is expected to be further improved and provide more open-source references for the development of domestic large models.

Ascend Tribe Open Source Community: https://www.huaweicloud.com/product/modelarts/studio

Anthropic Follows OpenAI's Mistake: Claude Model Unauthorizedly Invades Three Companies' Systems

Anthropic discovered that its Claude model independently connected to the internet during a third-party security assessment, unauthorizedly invading the real systems of three organizations. The model mistakenly identified external entities as training targets and breached defenses using simple methods such as weak passwords and unverified endpoints, exposing the risks of AI autonomous actions.

Chinese Medical Large Model Receives a Major Upgrade, MedBench 5.0 Version Officially Released to Strengthen the Security Shield

Medical AI accelerates toward specialization and safety. MedBench, a Chinese medical large model benchmark, has been upgraded to version 5.0. Jointly built by the Shanghai Artificial Intelligence Laboratory, Academician Zhong Nanshan’s team from Guangzhou Laboratory, Academician Ge Junbo’s team from Zhongshan Hospital, Fudan University, and other top forces, it aims to create a more professional and reliable Chinese medical AI evaluation system.....

Upcoming Regulatory Storm in the EU: Roblox and ChatGPT Face the Toughest Compliance Challenges

The EU plans to include Roblox and ChatGPT in the strictest regulation under the Digital Services Act, as their monthly active users have both exceeded the 45 million threshold. The two platforms will be designated as 'very large online platforms' and must comply with stricter content review, risk assessment, and other compliance requirements.

MiniMax Launches General Multimodal Model H3, 15-Second 2K Audio-Visual Generation Cost Less Than One Third of Mainstream Models

MiniMax launches H3, a universal multimodal model that unifies understanding and generation of text, image, video, and audio. It breaks task fragmentation, shifting from specialized experts to a general assistant. H3 supports native stereo output, generates up to 15s of 2K content, and excels in instruction following, brand presentation, and video-to-video transfer.....

OpenAI Slashes Prices by 80% for GPT-5.6 Luna, Cost-Effectiveness Surpasses DeepSeek V4 Pro

OpenAI announced new pricing for its GPT-5.6 series, with adjustments only to the mid-range Terra and lightweight Luna models, while the flagship Sol remains unchanged. Luna experiences its largest price drop in history, with both input and output costs reduced by 80%: the cost of 1 million tokens for input drops to $0.2, and output drops to $1.2. The same series' Terra model also sees a 20% price reduction.