News · 2026-10-16

Meta releases Muse Spark Open, but local use still demands enormous memory

Meta released Muse Spark Open on October 15 with downloadable weights and a published minimum of 700 GB of graphics memory for its NVFP4 serving configuration. The release expands access to a very large multimodal model, but its hardware needs and custom license determine who can actually operate it independently.

Key facts

• Meta describes 816 billion total parameters, with 42 billion active for each generated token.
• The release accepts text, images, video and audio, and produces text within a one-million-token context window.
• The technical walkthrough describes the BF16 checkpoint as about a 1.5 TB download; Meta specifies 2 TB of graphics memory for BF16 serving.
• Primary source: Meta’s October 15 announcement, with requirements and license in the official model card.

The most useful distinction is between obtaining the model and having the equipment to use it. An organization can apply through Hugging Face’s access gate and download the weights. That does not turn the release into a practical installation for a typical laptop or gaming computer. Meta’s own requirements place its official configurations in the territory of unusually large memory systems.

Muse Spark Open uses a mixture of experts: different parts of the network handle different inputs, and only selected parts perform work for a particular token. Meta’s card lists 256 experts, with eight activated per token. Think of an enormous reference library with a librarian opening eight relevant volumes for each question. Reading eight books saves work, but the library still needs somewhere to keep the entire collection.

That is why the 42-billion-active figure is a computation figure rather than a storage allowance. The Hugging Face technical walkthrough says all experts remain in memory. Meta separately specifies at least 700 GB of graphics memory for the low-precision NVFP4 configuration and 2 TB for BF16. File size and operating memory are different quantities: intermediate calculations, the conversation cache and runtime machinery also consume memory. The cache explainer describes one reason long conversations can become expensive even after the weights fit.

Meta also describes a desktop setup using two M5 Ultra Mac Studios, each supporting up to 512 GB of unified memory. This is a separate hardware route, documented in its platform guide. Unified memory should not be casually equated with a graphics card’s memory specification, and holding a model does not by itself establish an attractive response speed.

Third-party compression broadens the possibilities. The Unsloth repository lists a 176 GB one-bit variant and a 506 GB four-bit variant, among others. Those are published download sizes. They are not promises that a machine with exactly that much memory can serve an arbitrary conversation, and the dossier does not establish a full runtime requirement for those variants. Aggressive compression also calls for task-specific quality checks rather than assuming it preserves the original model’s behavior.

The legal details deserve equally close reading. The Muse Model License permits broad use, modification and distribution subject to its terms and usage policy. It requires the attribution “Built with Muse Spark Open.” For model-as-a-service providers whose aggregate gross revenue with affiliates exceeds $20 million in a consecutive twelve-month period, a separate agreement is required before offering the model or derivatives through that service. The reviewed license does not impose Llama’s familiar monthly-user cap.

Meta calls the release “the most capable US open-weight model to date.” That is the company’s characterization, not an independently established ranking. The dossier’s independent leaderboard check did not establish the previously circulated Artificial Analysis placement, so no rank belongs in the factual account. The launch discussion on Hacker News also shows why license interpretation matters alongside model performance.

The family relationship also needs care. Meta documents Muse Glimmer as a smaller model trained from Spark outputs through distillation. That does not make Spark Open a distilled Glimmer derivative. Nor does access to the new checkpoint establish an exact match to a closed service version: Meta explicitly describes additional safety post-training for the open release.

The practical gain is the ability to inspect, modify and host a very large model across supported runtimes, including the inference systems named in Meta’s release. The limitation is equally concrete: openness of weights does not guarantee affordable ownership, unrestricted commercial terms or identical behavior to a particular closed-service checkpoint. Buyers should evaluate the complete serving configuration and license for their intended use, rather than treating the active-parameter count as a purchasing specification.

Primary source, verified: read the paper →

Ground Truth. (2026, October 16). Meta releases Muse Spark Open, but local use still demands enormous memory. Ground Truth. https://groundtruth.day/news/muse-spark-open-releases-weights-with-a-700-gb-memory-floor.html

@misc{groundtruth:muse-spark-open-releases-weights-with-a-700-gb-memory-floor,
  title  = {Meta releases Muse Spark Open, but local use still demands enormous memory},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/muse-spark-open-releases-weights-with-a-700-gb-memory-floor.html}
}

Topics: meta · open-weights · mixture-of-experts · inference · licensing