The AI landscape shifted dramatically this week as Meta released Muse Spark Open, a sparse mixture-of-experts model with 816 billion total parameters, while NVIDIA shipped Nemotron 3.5 and OpenAI integrated ChatGPT directly into Chrome tabs. These developments signal a pivot toward massive, open-weight agentic models and deeper browser integration, all happening alongside Anthropic’s unprecedented move to grant outside auditors employee-level access to its internal systems.

Meta’s 816B Parameter Gambit

Meta’s new Muse Spark Open model activates roughly 42 billion parameters per token and supports text, image, video, and audio inputs with a 1M-token context window. The launch is notable for its broad day-one support across frameworks like SGLang, vLLM, and MLX, and hardware including NVIDIA, AMD, Intel, and Apple silicon. Alexandr Wang, Meta’s chief AI officer, framed the release as a bet that most AI will eventually run on open models from the US, with specific demos showing the model running locally on Mac Studio machines with M5 Ultra chips.

NVIDIA’s Open Agent Family

NVIDIA released Nemotron 3.5 Super VL, a 120-billion-parameter open multimodal model designed for always-on agents. It utilizes a hybrid Mamba-2, attention, and MoE architecture to achieve accuracy comparable to larger reasoning models with significantly higher throughput. The release includes Nemotron 3.5 Bolt 74B-A9B, which uses native 2-bit quantization to fit into 24GB of memory, enabling local agent execution on RTX and DGX Spark hardware. Major organizations like Scale AI, EY, and Ramp are already post-training the model for specific enterprise workflows.

Browser Wars and Governance

OpenAI announced that ChatGPT can now read and act within open Chrome tabs, directly competing with Google’s Gemini features in the browser. This move places the assistant platform on Google’s home turf, with OpenAI distinguishing between reading pages and performing actions like filling forms or sending messages. Meanwhile, Anthropic granted six evaluators from METR employee-level access to its internal tools, a governance step fulfilling Dario Amodei’s commitment to transparency after recent agent mishaps. Unlike previous attempts, Anthropic cannot veto METR’s publications, marking a significant shift in AI lab accountability.

Key Takeaways

  • Meta’s Muse Spark Open offers 816B parameters with 42B active, supporting multimodal agentic tasks across major hardware platforms.
  • NVIDIA’s Nemotron 3.5 family targets open agent workloads with efficient architectures and local deployment options.
  • OpenAI’s Chrome integration challenges Google’s dominance in the browser assistant space.
  • Anthropic’s partnership with METR introduces a no-veto public audit model for AI safety.

The Bottom Line

The industry is racing to commoditize intelligence via open weights while simultaneously tightening the noose of regulatory and public scrutiny. We are entering an era where the cost of inference drops, but the cost of trust rises, forcing labs to open their black boxes before the law forces them to.