06-reference

stratechery muse spark open meta apple thin thick

2026-10-20·reference·source: Stratechery·by Ben Thompson
stratecherymetaappleopen-weight-modelslocal-inferencethin-is-inagentic-codingchina-ai-models

Muse Spark Open, Meta + Apple, Thin and Thick (Stratechery Update, 2026-10-20)

Why this is in the vault

Three linked movements in one Update, all orbiting the same question: Meta ships an open-weight model (Muse Spark Open) co-engineered with Apple to run well on Mac Studio hardware, Thompson gets a hands-on demo of that setup on Apple's campus, and he uses it to re-stress-test his own "Thin Is In" thesis — the third time this thesis has come up for review in the vault this year.

The core argument

Muse Spark Open. Meta released a new open-weight model built for coding/agentic work, with Apple updating its MLX toolkit so it runs well on Macs. Like Meta's prior open releases, it carries restrictions beyond typical open source: mandatory "Built with Muse Spark Open" attribution, an acceptable-use policy, and a required commercial agreement once a reseller's revenue exceeds $20M/year. Some Chinese open models are genuinely permissive (DeepSeek-V4.1-Flash is MIT-licensed, smaller Qwen models are Apache), while others (Kimi K3, the largest Qwen models) carry similar restrictions to Meta's. Muse Spark Open is also simply weaker than top Chinese alternatives — it's based on a six-month-old checkpoint, and Thompson repeats his "Who's Afraid of Chinese Models" argument that Chinese labs benefit from distilling frontier models' outputs in ways US labs can't under their own terms of service, putting Meta at a structural quality disadvantage short-term. He ties this to the same week's DHS mandate ordering DeepSeek removed from federal civilian systems, arguing American defenders need access to the best available models (including, implicitly, ones restricted for cybersecurity use under current policy) or they're stuck defending with inferior tools against adversary-nation models.

Meta + Apple. Thompson was invited onto Apple's campus and saw two Mac Studios (512GB unified memory, linked via four Thunderbolt 5 cables) running Muse Spark Open locally — demoed building a mobile app and assembling a video clip compilation from raw TV footage. He reads the demo itself as secondary to the symbolism: an actual thaw between historically hostile Apple and Meta, which he attributes partly to John Ternus replacing Tim Cook as CEO. His sharper point is architectural: when local tokens are effectively free, "wasted" compute is costless, which opens a coding paradigm he compares to CPU branch prediction — generating multiple speculative solution paths overnight and picking the best one in the morning, rather than committing to one path at a time the way both humans and today's cloud-metered agents do.

Thin and Thick. Revisiting his February "Thin Is In" thesis (cloud wins on compute availability, memory, and shared-cost economics), Thompson treats the Mac Studio rig as the strongest real-world counterexample yet available — and concludes it still doesn't break the thesis. Muse Spark Open isn't frontier-grade, a comparably equipped Mac Studio setup runs toward $20K, and even in the demo a MacBook Pro acted as a thin client talking to the two Mac Studios as a personal server — still thin architecture, just with the "thick" part in the room. His conclusion: the economic case for local inference doesn't yet pencil out even against falling cloud API costs; the real case is "free as in speech" — independence from any cloud gatekeeper — which is a values argument, not a cost argument.

Mapping against Ray Data Co

The direct connection is architectural, not just thematic: Thompson's "branch prediction" coding pattern — generate several speculative build variants overnight because tokens are free, then pick one in the morning — is functionally the same shape as RDCO's own overnight delegation pattern (sw-builder at Opus-5.5-high-effort, "studio throughput, no deferral," up to 4 artifacts overnight for founder review in the morning). RDCO reached that pattern from a cloud-metered, not token-free, starting point, which means the economics Thompson treats as the open question (does free-token local compute change the shape of agentic work) are exactly the variable RDCO's current Anthropic-API-metered setup doesn't yet have — worth flagging as a structural dependency to watch, not act on, since Thompson's own verdict is that the thin-client, cloud-first architecture still wins on the numbers today.

Related