The day in AI Daily

The frontier is splitting into two strategies. Meta released the weights for Muse Spark Open, an 816-billion-parameter multimodal mixture-of-experts model with a one-million-token context window, while claiming it can run locally across two Apple M5 Ultra Mac Studios. That makes frontier-scale experimentation more accessible, although the practical cost, speed and licence constraints remain decisive. Meta AI Research

Microsoft, meanwhile, completed pre-training MAI-Base-2: a 2.8-trillion-parameter model with 80 billion active parameters. It is not yet generally available, and Microsoft’s performance and efficiency comparisons remain company-run claims, but the announcement confirms that hyperscalers are still investing heavily in in-house model pipelines rather than relying entirely on external providers. Microsoft AI

The deeper pattern

The most important development is not simply that the models are getting larger. It is that frontier capability is becoming a contest over the whole stack: model architecture, training data, specialised infrastructure, deployment hardware and independent evaluation.

Microsoft’s MAI-Base-2 illustrates the industrial end of this race. The company says it trained the model on 16,000 GB200 GPUs over 38 days, while improving mixture-of-experts efficiency and data pipelines. Its claimed 4.2–12.3-fold pre-training efficiency advantage over a reference open model should be treated as provisional until replicated, but the underlying direction is clear: compute efficiency is becoming as strategically important as raw parameter count. A model that is cheaper to train and serve can be deployed more widely, regardless of whether it wins a headline benchmark.

Meta’s Muse Spark Open points towards a different consequence: frontier-like models becoming more portable. Its weights are available through major hubs and inference providers, and Meta says the model can run on local unified-memory systems. That does not make a two-workstation setup ordinary consumer hardware, but it narrows the gap between “frontier model” and “model you can inspect, customise or host yourself”. Open weights therefore increase both experimentation and the difficulty of controlling downstream use.

The governance response is beginning to adapt to this reality. Anthropic says six METR employees now have employee-like access to internal tools and information, with publication rights that Anthropic cannot veto except for narrow redactions. Anthropic This is more substantive than commissioning an external benchmark, because evaluators can inspect parts of the development and safety process that outsiders normally cannot see. It is still not independent regulation: the arrangement is voluntary, small in scale and bounded by Anthropic’s security rules. Its value will depend on whether METR can publish uncomfortable findings and whether other providers accept comparable scrutiny.

What to watch next

  • Whether Meta publishes independent or reproducible evaluations showing Muse Spark Open’s claimed agentic and multimodal advantages, rather than relying mainly on vendor demonstrations.

  • Whether Microsoft releases MAI-Base-2 for Foundry or trusted partners, with model weights, pricing and post-training results. That would convert a training announcement into a competitive product consequence.

  • METR’s first substantive public assessment under its Anthropic arrangement: what access it received, what it tested, which limitations were imposed and whether its conclusions differ from Anthropic’s own risk reports.

Editorial note

The main blind spot is evidence asymmetry. The model claims come primarily from the companies that built them, while independent validation and real-world deployment data are not yet available.