Muse Spark Open

Original Article ↗ Hacker News Discussion ↗

Meta’s release of Muse Spark Open, a nearly trillion‑parameter “open‑weight” AI model, is praised as a significant step for locally run large models but criticized for its massive hardware demands—Meta showcases inference on clusters of high‑end Mac Studios—and for a license that restricts large “model‑as‑a‑service” providers. Commenters contrast Muse Spark Open’s performance and cost with leading Chinese open models like DeepSeek, GLM and Qwen, arguing that U.S. labs are increasingly marketing themselves as “best in the U.S.” rather than best overall. The launch also fuels broader debate over whether Meta can leverage its ecosystem to dominate consumer AI, how sticky agent-style products will become, and whether partially open releases meaningfully advance privacy and AI sovereignty.

Model nature, “open weights”, and licensing

  • Many are glad a major lab released weights, seeing it as strengthening the open‑weight ecosystem and raising the bar by providing deployment docs and tooling.
  • Others argue this is not truly “open”: license restricts large “model‑as‑a‑service” providers above ~$20M revenue, and the weights are gated via Hugging Face approval.
  • Debate over terminology: some insist “open‑weight” should imply an OSI‑style open‑source license; others say current industry usage just means “weights downloadable, with varying commercial terms.”
  • Some point out that copyright status of model weights is legally murky; companies and users will likely still treat licenses as binding.

Hardware and deployment realities

  • Official guidance targets two Mac Studio M5 Ultras with 512 GB each or four with 256 GB (≈1 TB unified memory), which many see as $25k–$50k territory.
  • Thread notes that 2‑bit quant still needs ~256 GB RAM/VRAM; average users will not run this interactively, only very slowly from SSD if at all.
  • Some appreciate explicit optimization for Mac/MLX and broad inference‑engine support (Unsloth, etc.), but others call it marketing lip service given the massive requirements.

Performance vs other open models

  • Benchmarks and user tests suggest Muse Spark Open is:
    • Below Muse Spark 1.3, closer to 1.2 on several coding benches.
    • Roughly similar to DeepSeek V4.1 Flash and Mimo 2.5 Flash on Artificial Analysis, but much more expensive to run.
    • Outperformed by GLM 5.3 Flash and Qwen 3.8 Flash Next in intelligence‑per‑cost and hardware practicality.
  • Some image‑to‑HTML and Pelican SVG tests show it’s “solid” but not standout; Muse Wonder and frontier closed models still look stronger.

Meta’s strategy and “consumer AI”

  • One camp sees Meta executing a deliberate moat strategy: open big models, make them easy to deploy, drive ecosystem standardization around its stack, and funnel users into its cloud “Muse” agent.
  • Others doubt “stickiness”: personal agents lack strong network effects; UX alone is not a durable moat; top Chinese and frontier models remain more capable.
  • Some think Meta, Google, and others are carving out niches (“best US/EU open weights” vs Chinese models), appealing to companies unwilling to use Chinese tech.

Trust, safety, and geopolitics

  • Several express distrust of Meta’s data practices and “open‑washing.”
  • Others are more worried about Chinese open models, citing potential backdoors or subtle sabotage, while critics respond that Western actors have also backdoored tech historically.
  • Bio/chem misuse risk is debated: some argue powerful open models are dangerous; others note that basic chemical/biological weapon knowledge is already widely available in textbooks.