rbtfl.
Meta's Superintelligence Labs raids Thinking Machines, goes closed

Meta's Superintelligence Labs raids Thinking Machines, goes closed

Five Thinking Machines founders hired, one on a reported $1.5bn package; Muse Spark breaks with the open-source Llama tradition

AI·open-vs-closed· transition 誰の金か·静かな変化 ·6 論調 · ·rbtfl 更新 2026年6月25日

Summary

Meta AI's Superintelligence Labs, built around the $14.3bn Scale AI stake and led by Alexandr Wang with Nat Friedman, escalated the talent war by hiring five founding members of Mira Murati's Thinking Machines Lab after Murati rejected a ~$1bn acquisition offer. Co-founder Andrew Tulloch reportedly received a $1.5bn-over-six-years package; Meta disputes the figure as stock-contingent. Not all defected to Meta, Barret Zoph, Luke Metz and Sam Schoenholz returned to Openai. The lab's first output, Muse Spark (8 April), is a closed multimodal reasoning model now powering Meta AI across its apps and Ray-Ban glasses, a deliberate break with the open-weight Llama tradition.

By the numbers

  • $14.3bn, Meta's 49% non-voting Scale AI stake.
  • 5, Thinking Machines founders hired by Meta.
  • ~$1.5bn / 6 years, reported Tulloch package (Meta disputes).
  • 8 Apr 2026, Muse Spark released (closed-source).

Why it matters

Meta is converting cash into scarce researchers at unprecedented prices and walking away from open weights at the top, a strategic reversal that thins the open-source frontier and signals the IP from these hires stays proprietary. The bidding war redraws comp benchmarks across every lab.

What to watch

  • Whether Llama continues as an open line or is fully superseded by closed Muse models.
  • Retention: how many expensive hires stay through vesting.
  • Whether the closed pivot dents Meta's open-ecosystem goodwill.