The artificial intelligence landscape continues to bifurcate and accelerate as two tech giants drop major architectural updates. According to newsletters from AI Breakfast and TLDR AI, Meta Superintelligence Labs has released Muse Spark Open, a massive 816 billion-parameter open-weight model tailored for local execution on Mac hardware. Simultaneously, Microsoft pulled back the curtain on its proprietary MAI-Base-2 training stack, a 2.8 trillion-parameter mixture-of-experts model designed to power its enterprise offerings.
Meta's Muse Spark Open enters the market as a formidable contender in the open-weight space. Featuring an 816 billion-parameter multimodal architecture and a sprawling one million-token context window, Meta has positioned the model as the most capable US open-weight release to date. Crucially, the model is engineered for local execution across Mac hardware, signaling a continued push to bring high-end frontier capabilities closer to edge devices and local developer environments.
On the proprietary front, Microsoft detailed the engineering behind MAI-Base-2 in a post from Microsoft AI. Built following the June release of MAI-Thinking-1, MAI-Base-2 is a mixture-of-experts foundation model boasting 2.8 trillion total parameters and 80 billion active parameters. Microsoft trained the model across 16,000 GB200 GPUs over a grueling 38-day period in August and September 2026, achieving a model FLOPs utilization of 21 percent and a goodput of 89 percent.
According to Microsoft, MAI-Base-2 outperforms every tested open-weight model on held-out evaluations and is four to 12 times more pre-training FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base. Microsoft emphasized that its data provenance relies entirely on organic data without distillation from other models, allowing the company to license model weights to trusted enterprise partners through Foundry.
For founders and business leaders, these simultaneous releases highlight two distinct strategic paths. Meta's open approach empowers builders to harness massive local capabilities without heavy cloud dependencies, lowering the barrier for sophisticated local application deployment. Meanwhile, Microsoft's tightly controlled, end-to-end stack demonstrates the immense efficiency gains of co-designing hardware and training infrastructure, offering enterprises a compliant, audited path to frontier intelligence.