Apple ML Research: Muse Spark Open with MLX on Mac Studio

Original Article ↗ Hacker News Discussion ↗

Apple’s latest ML research post showcases how its Mac Studio M-series desktops can run Meta’s near-trillion-parameter Muse Spark Open model locally using the MLX framework and Thunderbolt-based clustering. Commenters see this as an impressive proof of concept but question the practicality, noting that multi–Mac Studio setups cost tens of thousands of dollars compared with relatively cheap API access to frontier models. The exchange highlights broader tensions over local vs. cloud inference, Apple’s emerging role as a high-end “shovel seller” for AI workloads, and whether open-weight models are progressing toward efficient, truly commodity hardware deployment.

Cost and economics of local inference

  • Running Muse Spark Open on four high‑end Mac Studios is seen as impressive but very expensive (tens of thousands of dollars).
  • Several commenters compare this to paying for API/ subscription access (e.g., via OpenRouter) and argue that, for many, credits are cheaper and more flexible given rapid hardware/model churn.
  • Others note personal experiences where owning hardware paid off over many years vs renting, stressing that the tradeoff depends heavily on utilization.

Model choice, benchmarks, and performance

  • Some think Apple sandbagged alternatives by showing Muse Spark Open at 4‑bit while e.g. Qwen is only shown in BF16, which isn’t how users would typically run it.
  • Multiple commenters claim Muse Spark underperforms smaller/open models (e.g., 300B‑class) on quality benchmarks, though it is praised for speed and token efficiency.
  • Others would rather run models like GLM 5.3 Flash, Qwen, or DeepSeek on the same hardware.

Apple–Meta relationship and marketing angle

  • Several people find it odd Apple is showcasing a Meta model, given their past tensions.
  • Some frame this as cross‑promotion: Meta spends to build open‑weight models; Apple sells expensive hardware to run them.
  • The official Apple article title focuses on MLX/RDMA, but the thread notes the Muse emphasis feels like co‑marketing.

Hardware strategy and AI ecosystem

  • Many argue Apple is “selling the shovels” rather than needing to win the model race.
  • There’s debate over whether Apple is behind in AI infrastructure (no CUDA/TPU‑like stack, no real datacenter offering) versus simply playing a different game focused on consumer devices and vertical integration.
  • Some see this as an early step toward widespread on‑device “good enough” inference that could hurt Nvidia long‑term.

Privacy, Photos, and local models

  • One camp sees local models and Apple’s Photos/“Foundation” models as privacy‑oriented and useful for on‑device search/filtering without sending data to cloud LLMs.
  • Another camp disputes that iCloud Photos is genuinely private, even with advanced options, and argues that default settings and metadata undermine the privacy story.

macOS background load and user experience

  • A subthread reports heavy, constant CPU use from system daemons (media analysis, iCloud/phone sync, Spotlight, etc.), allegedly worsened by new “AI” features and causing heat/throttling on M‑series Macs.
  • Others say their experience is the opposite (machines stay cool and fast), suggesting possible bugs or edge cases; the true prevalence is unclear.