Meta launched Muse Spark, its first frontier model under Meta Intelligence Labs, beating rivals on some benchmarks while abandoning its open-source stance.
Meta announced Muse Spark, its first major model release since restructuring its AI division as Meta Intelligence Labs, led by Scale AI CEO Alexandr Wang following a $14.3B investment. The model is available via meta.ai and the Meta AI app, and self-reported benchmarks claim it outperforms models from OpenAI, Anthropic, Google, and xAI on some tasks. Critically, unlike the Llama series, Muse Spark is closed source — a major strategic reversal for a company previously seen as the open-source AI champion. Meta published an Advanced AI Scaling Framework alongside the release, outlining safety checks as models approach superhuman performance.
Developers who built on Llama's open weights for fine-tuning, local deployment, or cost-optimized inference now face a dead end with Muse Spark. The model is API/app-only, which means no custom training, no self-hosting, and no escaping per-token pricing. Meta promises future open-source releases, but that's a roadmap promise, not a shipping date — your architecture decisions need to be made against what's available today.
Run Muse Spark via meta.ai on your hardest benchmark prompt this week and compare latency and output quality against your current Llama or GPT-4o integration — if it wins, evaluate whether the closed-source API tradeoff is worth the switch before Meta's pricing is formalized.
Go to meta.ai and open a new chat with Muse Spark
Tags
Also today
Signals by role
Also today
Tools mentioned