Artificial intelligence news.
Straight from the original announcements.

Meta's Muse Spark Open arrives on Hugging Face as an open-weight MoE model

October 15, 2026By the AI News Bites Editorial TeamSource: Hugging Face
Top storyMeta's Muse Spark Open arrives on Hugging Face as an open-weight MoE model
Image: Hugging Face

Hugging Face has welcomed Muse Spark Open, a model from Meta that the platform describes as an open-weight, frontier, multimodal mixture-of-experts model. The announcement appeared on the Hugging Face blog on 15 October 2026.

Key points
  • Hugging Face published a blog post welcoming Muse Spark Open from Meta on 15 October 2026.
  • The model is described as open-weight, frontier, multimodal and mixture-of-experts.
  • The page mentions one paper, arXiv 2504.13181, published on 17 April 2025, and one dataset.
  • The available text gives no size, licence, benchmark or availability details.

What Hugging Face published

The Hugging Face blog post is titled "Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta". The title is the clearest statement of what the release is. It names the model, Muse Spark Open, and credits Meta as its source.

The page text available for this report is limited. It lists one dataset and one paper mentioned in the article. The paper is arXiv 2504.13181, published on 17 April 2025. The excerpt does not say how the paper relates to the model, so that link should not be assumed.

What the description means

Each term in the title points to a specific kind of model.

  • Open-weight: the trained model weights are made available, which generally lets developers run and study the model themselves rather than only reaching it through a hosted service.
  • Multimodal: the model is built to work with more than one type of data, not only text.
  • Mixture-of-experts: an architecture in which a model is split into specialised sub-networks, and only part of the model is used for each input.
  • Frontier: the label Hugging Face uses to place the model among the most capable systems available.

Taken together, the title positions Muse Spark Open as a large-scale model from a major AI developer that the wider community can access directly.

What is still unclear

The text provided does not include the details that usually decide how useful a release is. It gives no parameter count, no licence terms, no benchmark results, no supported input types and no hardware requirements. It also does not describe how the model can be downloaded or deployed. Any claims about performance or availability will need to come from the full Hugging Face post or from Meta.

The page also lists two other recent Hugging Face blog posts. One is about Jun Kim, the creator and maintainer of oMLX, joining Hugging Face to support the MLX community, dated 22 September 2026. The other is titled "DeepSeek-V4: a million-token context that agents can actually use", dated 24 April 2026. Both show that Hugging Face has been covering open model releases and tooling in recent months.

For developers and researchers, the practical significance of Muse Spark Open will depend on those missing details. What the title does establish is that Meta's model is being presented as open-weight and published through Hugging Face, one of the main hubs for sharing open models.

Related news

Products and tools
Products and toolsNovember 12, 2026, 20:31

Snowflake Adds Inkling 2 Small to Cortex Training

Snowflake has made Inkling 2 Small available in Cortex Training public preview, giving teams a way to benchmark, post-train, evaluate and serve the open-weight model on…

Vercel Adds Meta’s Muse Ember to AI Gateway for Agent Tasks
Products and toolsNovember 9, 2026

Vercel Adds Meta’s Muse Ember to AI Gateway for Agent Tasks

Vercel has made Meta’s Muse Ember available through AI Gateway, positioning the lightweight reasoning model for repeated, bounded tasks inside agent runs. The model…

Baseten launches joey-1, a 2B decision model with 20 ms latency
ModelsOctober 29, 2026, 18:24

Baseten launches joey-1, a 2B decision model with 20 ms latency

Baseten has released joey-1, a 2B-parameter decision model that, according to the company, serves an average 2,000-token request at a P50 end-to-end latency of 20 ms. It…