Trillion-Scale Race, Free Frontier, and Safety's C+ Ceiling

A week of extremes: Chinese labs disclosed the largest training runs in their history, OpenAI removed the paywall from its frontier model tier, and an independent safety audit found no developer worthy of more than a passing grade.

ByteDance Trains China's Largest AI Model

TikTok's parent company, ByteDance, is pre-training an artificial intelligence model with up to 10 trillion parameters — roughly three times larger than Moonshot AI's Kimi K3 and approaching industry estimates for Anthropic's most advanced Mythos system. The model relies on a Mixture of Experts architecture, which routes each input to only a small fraction of the total parameters, making such extreme scale technically and economically viable. Pre-training typically spans three to six months before fine-tuning can begin; no release date has been announced. The disclosure confirms that China's largest technology companies are no longer watching the frontier from a distance.

Alibaba Open-Sources Qwen3.8-Max

On August 3, Alibaba released Qwen3.8-Max, a 2.4-trillion parameter Mixture of Experts model that activates 95 billion parameters per inference. The model handles text, images, and video with a one-million token context window, and in internal benchmarks Alibaba says it performs on par with — and sometimes better than — Anthropic's Fable 5. A demonstration showed the model completing a multi-day software engineering project from scratch without human intervention. Alibaba shares rose 7 percent in Hong Kong on the announcement. The company has committed to releasing the model's open weights, continuing the strategy that made the Qwen family widely adopted outside China.

OpenAI Brings GPT-5.6 Luna to All Free Users

OpenAI made GPT-5.6 Luna the default model for free and entry-level ChatGPT accounts this week — the first time a member of the GPT-5.6 family has been available without a paid subscription. Luna outperforms GPT-5.5 Instant on every benchmark OpenAI has published, despite being the smallest member of the new family. Starting next week, free users will receive unlimited text conversations and a "Think" button for problems that benefit from deeper reasoning. The move follows an 80 percent API price cut for Luna on July 30. MacRumors noted it is the broadest expansion of frontier AI access in ChatGPT's history.

NVIDIA Open-Sources NOOA: Agents as Python Classes

NVIDIA published version 0.0.8 of NOOA (Object-Oriented Agents) on August 7 — an Apache 2.0 Python framework that represents an AI agent as a single class, with fields for state, methods for capabilities, and docstrings as prompts. On the SWE-Bench Verified coding benchmark, a NOOA-based configuration achieved 82.2 percent accuracy using GPT-5.5, beating the previous record of 79.2 percent while cutting token usage by up to 50 percent. The library is model-agnostic and installs with pip install nooa. NVIDIA's developer blog describes six harness capabilities the framework adds to any agent loop.

No AI Lab Earns Above C+ on Safety

The Future of Life Institute's Summer 2026 AI Safety Index graded nine leading AI developers across six domains — transparency, safety frameworks, risk assessment, technical research, governance, and societal accountability. None scored higher than a C+. Anthropic led the field at 2.66 out of 4.0, followed by OpenAI (C, 2.28) and Google DeepMind (C, 2.01). Meta received a D+, and three labs — xAI, DeepSeek, and Mistral — received outright F grades. Researchers noted that commercial safety progress has largely plateaued, and several leading labs have scaled back safety commitments to pursue defense contracts. The index offers a sober reminder that rapid capability growth has not been matched by governance progress.

Cyprus Opens AI Strategy 2032 Consultation

Cyprus is inviting public input on its National AI Strategy 2032, a plan built on the principle that "AI must serve people." The strategy aims to position the island as a trusted regional AI hub in the Eastern Mediterranean, with specific commitments to reach small businesses, people with disabilities, and rural communities. The consultation runs until August 31. In parallel, high-risk AI requirements under the EU AI Act took effect on August 2 and now apply to businesses and public bodies across Cyprus. The Deputy Ministry of Research, Innovation and Digital Policy is preparing guidance for small enterprises and plans to open a national AI regulatory sandbox. Business AI adoption on the island remains below the EU average, making the timely implementation of this strategy particularly consequential.