Model News
Huawei opens 505B Pangu weights trained on Ascend, not Nvidia
By Arjun
·
7 Aug 12:04 AM
·
Via Huawei / Hugging Face
Huawei posted weights and inference code for openPangu-2.0-Pro, a roughly 505 billion parameter mixture of experts model with about 18 billion active parameters per token and a 512,000 token context. The lab says pretraining used Ascend 910B NPUs and about 34 trillion tokens, with no Nvidia GPUs in the cluster. Checkpoints are on Hugging Face under Huawei's OpenPangu license, which restricts EU use. Why it matters: a frontier scale open weight stack on Chinese silicon is now downloadable. Caveat: Ascend parts have previously included foreign dies and memory, so Nvidia free training is not a fully domestic supply chain, and independent benchmarks remain thin.
Read the original →
Our summary is original writing; the full story belongs to Huawei / Hugging Face.
Sign in or create a free account to keep this story in a bookmark folder.
Tencent said Hy3's production model, released in July 2026, now ranks among the top three models globally by OpenRouter token use, while WorkBuddy and CodeBuddy lead AI office and coding tools inside China. Second-quarter capital expenditure hit RMB 52.8 billion, up 176 percent year on year, mainly for AI compute behind Hy upgrades, WorkBuddy and CodeBuddy inference, Weixin AI, and cloud demand. Free cash flow turned negative after capital expenditure payments exceeded operating cash flow. Management frames Hy3 as a step toward later Hy models. Capex intensity is clear in the filing, yet Hy4 is only later this year, so the next leap is unproven.
Alibaba released downloadable weights for flagship Qwen3.8-Max while adding commercial terms that force large monetizing users to buy a separate Qwen licence. South China Morning Post reports the rule hits affiliates running model-as-a-service or AI work-assistant businesses with more than $50 million in aggregate revenue over any consecutive 12 months. Internal use stays exempt if the model, outputs, or capabilities are not offered to third parties. Most developers can still copy, modify, and distribute the 2.4-trillion-parameter sparse MoE for free if they pay their own compute. How Qwen will detect threshold breaches, and how derivatives are defined, remains unclear outside the licence text.
SpaceXAI launched an early beta of Grok Bot as always-on AI teammates that keep a shared cloud computer, sign into apps and websites, and finish multi-step jobs until they need approval. Users can run several bots in parallel, including a coordinator that assigns work, and message them from desktop or iOS. Access starts for SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium subscribers, with enterprise users on a waitlist. The product lands after SpaceX's agreed Cursor acquisition and amid workplace agents from OpenAI, Anthropic, and Microsoft. Reliability across arbitrary apps, and how credentials are shared among bots, still need real-world scrutiny in gated beta.
AI coding startup Cursor plans to open its first India office by year-end, Asia-Pacific head Simon Green told Moneycontrol, after hiring remotely across Bengaluru, Delhi, Mumbai, Hyderabad, and Chennai. India is now its third-largest market, with more than 3 million developers after tripling in a year. Cursor launched India-priced Start at Rs 649 a month with UPI, between Free and the roughly Rs 2,000 Pro tier. The move lands while SpaceX's agreed acquisition is still expected to close in the third quarter. Office timing and headcount were not spelled out, so this remains a market signal more than a completed build-out.