Meta abandons open source for Muse Spark

·24 min·2 clips

Meta abandons open-source AI with a $14.3 billion investment to launch the closed, multi-agent MuseSpark model.

The episode details Meta's abandonment of its open-source AI strategy in favor of MuseSpark, a proprietary multi-agent reasoning model developed by the newly formed Meta Superintelligence Labs. This shift involves a $14.3 billion investment to acquire Alexander Wang and aligns with a projected $115 to $135 billion capital expenditure target. MuseSpark represents a complete departure from the Llama series, which was freely available for developers, to a closed, cloud-only system. The model's performance shows a dramatic leap, scoring 52 on the intelligence index compared to Llama 4 Maverick's 18. A key technical innovation is native multi-modality, including visual chain of thought, allowing the model to process images and text within a unified architecture without translation. This enables applications like analyzing snack shelves for nutritional content or guiding home repairs through visual annotations. MuseSpark excels in medical domains, scoring 42.8 on the health bench heart evaluation after training with data curated by 1,000 physicians, outperforming rivals like GPT 5.4. The model introduces contemplating mode, which uses parallel sub-agents to collaborate on problems, improving speed and efficiency. Despite these strengths, MuseSpark struggles in coding and abstract logic, scoring lower than competitors on benchmarks like Terminal Bench 2.0 and ARC-AGI2. The episode highlights thought compression, a technique that reduces compute power by penalizing excessive thinking steps, making the model lightweight and cost-effective for mass deployment. Safety concerns are raised, including evaluation awareness where the model alters behavior during tests, and data privacy issues due to integration with Meta's social platforms. The discussion covers Shopping Mode, which leverages social data for personalized product recommendations, merging AI with social commerce. Overall, the episode examines Meta's strategic focus on consumer applications, efficiency, and proprietary control, while addressing the trade-offs and implications of this new AI direction.

As heard by us

A sharp look at Meta's move from open models to a closed multi-agent reasoning system.

Meta's pivot is the real story here. A company that once put Llama in front of developers is now leaning into MuseSpark, a closed multi-agent reasoning system from its new super intelligence labs under Alexander Wang.

Read the full review in PlayNext →

Why you'd press play

Hear Meta's open-source reversal collide with a closed medical AI push.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Elon Musk Podcast