Muse Spark: Scaling towards personal superintelligence

Original Article ↗ Hacker News Discussion ↗

Access & Availability

  • Muse Spark is currently only accessible via meta.ai and Meta apps (Facebook, Instagram, WhatsApp), not via public API or open weights.
  • There is a “private preview” API for selected partners; details on who qualifies and when broader access arrives are unclear.
  • Many commenters want a simple self‑serve, pay‑as‑you‑go API model; others say Meta mainly built this to embed across its own properties, not as a general developer platform.

Performance, Benchmarks & “Benchmaxxing”

  • Meta’s own benchmarks place Spark roughly in the frontier tier, sometimes close to or slightly ahead of other leading models, but:
    • Several commenters highlight that earlier Llama 4 benchmarks were misleading (“benchmaxxed”), making them skeptical of Meta’s numbers now.
    • Some point to weak scores on reasoning benchmarks (e.g., ARC-AGI v2) and lagging behind the latest Anthropic models on hard reasoning tasks.
    • A few early testers report basic math and analytical errors, saying it feels below GPT/Gemini/Claude in reliability; others report surprisingly strong results on specific tasks.
  • Consensus: promising but not clearly SOTA; claims need independent evaluation.

Multimodality & Use Cases

  • Commenters see visual reasoning and multimodal capabilities as the most impressive aspect; some report it outperforming other top models on complex document/floor‑plan tasks.
  • Many expect its primary value to be powering Meta’s consumer products (Marketplace, messaging, small‑business tools) rather than being a preferred standalone coding or research model.

Open Source, Ecosystem & Strategy

  • Thread repeatedly asks whether Meta has abandoned open‑weight releases; official language only “hopes” to open‑source future versions.
  • Some argue Meta previously accelerated the entire open ecosystem with Llama and has now lost that strategic and reputational advantage.
  • Others note that even being “4th place” still matters internally: cost control, independence from OpenAI/Anthropic/Google, and long‑term platform control.

Privacy, Trust & UX

  • Strong concern about Meta using chats to train models and its broader data‑harvesting reputation; several commenters refuse to try Spark for that reason.
  • Login is required (FB/Instagram), with reports of broken authentication flows and dark‑pattern UX (typing a prompt then being forced to log in).
  • Mixed sentiment overall: technical curiosity and appreciation for more competition, tempered by distrust of Meta, skepticism about hype (“personal superintelligence”), and frustration over closed weights.