Meta puts a frontier-class open model on two Mac Studios
Meta has released the weights for Muse Spark Open, which it calls the most capable US open-weight model to date. It is natively multimodal, with 816B total parameters (42B active per token) and a 1M-token context window, and it is tuned for coding and long-horizon agent work. The headline trick is local: with Apple's MLX framework it runs across two M5 Ultra Mac Studios, whose combined terabyte of unified memory holds the whole model. It is also live on Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together, with weights on Hugging Face and the big clouds. The implication: serious agentic coding no longer strictly needs someone else's data centre, and US labs now have a real answer to China's open-weight releases.
What else matters
EU sends its first AI Act information requests
Brussels has sent its first formal AI Act information requests to OpenAI and Anthropic, tied to this summer's agent security incidents during testing. They ask about security, external evaluation and usage monitoring, and they are not findings or fines. Still, it marks the Act's shift from rulebook to live enforcement.
ASML says it can't build machines fast enough
ASML reported record Q3 sales of €11.7 billion and warned that its planned capacity for 2027 and 2028 won't fully meet demand, with some customers already discussing orders for 2029 and 2030. For the AI buildout, lithography supply is becoming the next bottleneck after power.
Nvidia commits $250M of compute to UK AI safety research
Nvidia will provide $250 million in compute over five years for UK research into safeguards and control of AI agents, with Oxford and UCL among the groups involved. Jensen Huang also said, hypothetically, he would stop supplying chips to a developer building unsafe AI.
