Meituan Reveals 1.6 Trillion Parameter Model Trained on Chinese Chips
On June 30, Chinese delivery giant Meituan unveiled its next-generation open-source model, "LongCat-2.0." The company announced that it has transitioned to open-source and claims that this model is the world's first AI system with 1.6 trillion parameters, fully trained and operated on a cluster of 50,000 chips made entirely from domestic processors.
The model features a context window of 1 million tokens, comparable in scale to DeepMind's latest flagship model, V4-pro, released in April. Notably, it was trained without NVIDIA GPUs, utilizing Huawei's HCCL-based infrastructure, marking a significant example of China's "de-NVIDIA" efforts amid prolonged U.S. export controls.
Meta Tested Competitors' Chatbots Using Child Impersonation
According to reports citing Wired, hundreds of contract workers involved in a Meta project were instructed to impersonate minors to input sensitive prompts related to suicide, sexual topics, and drugs into competitors' chatbots, including Google Gemini and OpenAI's ChatGPT.
The project, named "Cannes," was operated by contractor Covalent, which reportedly submitted over 45,000 prompts in just one round in August 2025. However, Meta responded to Wired, stating that testing chatbot responses to ensure safety and age-appropriateness is a responsible industry standard and denied using benchmarking results from competitors for training its AI models.
Anthropic's Claude Mythos 5 Partially Reopens Amid U.S. Export Controls
On June 29, Anthropic's Claude Mythos 5 was partially reopened to core infrastructure defense companies under U.S. export control guidelines, while access for general users remains suspended for the 18th day. Key dates ahead include July 8, when Anthropic's revised privacy policy, which includes government ID verification, takes effect, and August 1, the deadline for the U.S. National Security Agency, Treasury Department, and Cybersecurity and Infrastructure Security Agency (CISA) to establish a review system for frontier models. In the meantime, Claude Opus 4.8 is serving as the basic model for API and Claude code.
GitHub Copilot Faces Developer Cost Shock After Switch to Usage-Based Billing
On June 30, marking one month since GitHub Copilot transitioned from a flat-rate subscription to a usage-based "AI credit" system, developers reported significant increases in costs. While the base subscription fee remains unchanged, new credit limits have been set. Reports indicate that monthly fees for using agent workflows have surged from $29 to $750 and from $50 to $3,000. This shift aligns with a trend of more detailed management of corporate AI spending, with Goldman Sachs projecting a 24-fold increase in token consumption by AI agents over the next four years.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.