Model Trained

• On August 2, 2026, the European Union's landmark AI law will fully enter into force, granting the Commission the authority to fine developers of powerful AI models. • Providers of models classified as having "systemic risk" must adhere to strict mandates, including rigorous stress-testing, security maintenance, and the mitigation of risks to public health and safety. • These regulations shift the AI Act from a set of deadlines to an enforceable legal framework, ensuring that the most influential AI systems are monitored for societal and security impacts.

• Chinese food delivery giant Meituan has unveiled LongCat-2.0, a new open-source large language model featuring 1.6 trillion parameters. • The company claims this is the first model of its scale to complete both full training and inference using exclusively domestic Chinese chip clusters. • This development is significant as it demonstrates China's ability to build massive AI models despite international restrictions on high-end semiconductors.

• Meituan unveiled LongCat-2.0, a new large language model that the company claims performs comparably to Google's Gemini 3.1 Pro released in February. • The model was trained using domestic Chinese chips, marking a significant step in the country's effort to reduce reliance on U.S. hardware. • This development is a direct response to Washington's security-based restrictions on advanced chip exports to China.

• Researchers at Sapient developed a 1B parameter reasoning model, called HRM-Text, by training it from scratch for approximately $1,500. • The model was trained on 40B tokens and achieved performance levels competitive with larger models ranging from 2B to 7B parameters. • This breakthrough demonstrates that foundational pretraining is no longer exclusive to wealthy institutions, allowing smaller organizations to build capable reasoning models affordably.