LaunchpadHQ
All model families

Meta

Muse (Meta) models

Version comparison

Last verified Oct 24, 2026. Figures that could not be confirmed on an official vendor page are marked unverified — we do not guess.

Full Muse (Meta) comparison

Which one should I use?

  • Meta's open models are now the Muse family; llama.com leads with Muse. Muse Spark is the model named in the consumer Meta AI assistant, and Muse Spark Open is its downloadable open-weight release. We don't treat them as the same checkpoint.
  • Pick Muse Spark Open for a frontier open-weight model if you have data-center hardware or use a hosted provider. Pick Muse Glimmer 30B if you want one that runs on a single high-end consumer GPU.
  • The licenses differ: Muse Spark Open uses Meta's Muse Model License, while Muse Glimmer 30B is Apache 2.0.

Muse Spark

Current
Context
Unverifiedunverified
Max output
Unverifiedunverified
API input / 1M
Unverifiedunverified
API output / 1M
Unverifiedunverified

Best for: People using the consumer Meta AI assistant, not self-hosting weights.

Named on Meta's consumer assistant page. Specs for that consumer listing were not published on the Muse Spark Open docs, and we do not treat the two as the same model.

Source: Meta AI assistant

Muse Spark Open

Current
Context
1.05M
Max output
Unverifiedunverified
API input / 1M
Unverifiedunverified
API output / 1M
Unverifiedunverified

Best for: Teams that want Meta's downloadable open-weight Muse Spark release or a third-party hosted API — not the consumer assistant app.

Downloadable open-weight release (also via hosted APIs). Not a model of the consumer Meta AI assistant app. Meta's blog describes running on two Apple M5 Ultra Mac Studios with MLX.

Source: Muse Spark Open docs

Muse Glimmer 30B

Current
Context
131.1K
Max output
Unverifiedunverified
API input / 1M
Unverifiedunverified
API output / 1M
Unverifiedunverified

Best for: People who want a capable Meta model running privately on their own computer.

Free download on Hugging Face (not gated). The model card says it runs locally in about 24–32 GB of memory with 4-bit versions.

Source: Muse Glimmer 30B model card

Muse Spark

People using the consumer Meta AI assistant, not self-hosting weights.

  • Consumer-assistant listing on Meta's official assistant page.

Caveats

  • Keep this distinct from Muse Spark Open (the downloadable open-weight release).
  • Context, output limits, and API token prices are unverified for this listing.

Hosted API prices are set by providers, not Meta. No first-party per-1M-token list price — we do not publish provider rates.

Muse Spark Open

Teams that want Meta's downloadable open-weight Muse Spark release or a third-party hosted API — not the consumer assistant app.

  • Open-weight release of Muse Spark (15 Oct 2026).
  • 816B total / 42B active sparse mixture of experts.
  • 1,048,576-token context; text, image, audio, and video input; text output.
  • Checkpoints: BF16 ~1.65 TB (8×B300); NVFP4 ~537 GB (4-GPU Blackwell node). Budget roughly 2 TB VRAM for BF16.
  • Weights via Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.
  • Hosted API via OpenRouter, Together AI, Fireworks AI, Groq, Baseten, Modal, and Nebius.

Caveats

  • Muse Model License: free to use, modify and sell, but you must show “Built with Muse Spark Open”, and model-as-a-service businesses over $20M in yearly revenue need a separate agreement with Meta. Not intended for anyone under 18.
  • Hosted API prices are provider-set; no pricing is published here.
  • Max output tokens unverified on the official pages checked.

Hosted API prices are set by providers, not Meta. No first-party per-1M-token list price — we do not publish provider rates.

Muse Glimmer 30B

People who want a capable Meta model running privately on their own computer.

  • Open weights released August 2026.
  • About 30B-parameter dense model with a vision encoder: text and image in, text out.
  • Built for agent-style tasks on consumer hardware; distilled from Muse Spark.
  • Comes in BF16, GGUF and ExecuTorch versions, plus a small drafter model for faster output.

Caveats

  • Much smaller than Muse Spark Open, so expect weaker results on the hardest tasks.
  • The model card says it isn't intended for anyone under 18.

Hosted API prices are set by providers, not Meta. No first-party per-1M-token list price — we do not publish provider rates.