Skip to content
AI Radar by Wagner Solution
CS
Menu
important New models verified update

Muse Spark Open weights are available on Hugging Face, with conditions set by the license

clearly official source updated 1 h ago

The weights for Muse Spark Open from Meta are available on Hugging Face. The free license requires the attribution “Built with Muse Spark Open”; API providers with annual revenue above 20 million dollars need a separate agreement.

The weights for Muse Spark Open are available on Hugging Face. The license allows free use but requires the attribution “Built with Muse Spark Open”. Providers with annual revenue above 20 million dollars that want to offer the model as a service via an API need a separate agreement with Meta.

Meta designed the model for programming and agentic tasks, such as reading through a project, editing and debugging code, or calling tools. The model accepts text, images, audio and video and produces text output. The context window holds up to a million tokens across all input modalities combined. The model has approximately 816 billion parameters, of which about 42 billion are active for each token; eight of 256 experts are selected in each MoE layer.

According to Meta, the model is optimized for the M5 Ultra chip from Apple and can run on two Mac Studio computers with 512 GB of unified memory each. The weights are available in BF16 and NVFP4 formats, with the BF16 variant taking up approximately 1.5 TB. The Hugging Face blog also describes running the BF16 variant on four nodes with eight H200 GPUs each, and the four-bit NVFP4 variant on a single B200 node; this variant is intended for NVIDIA Blackwell GPUs.

What changed

Why it matters

The available weights make it possible to deploy the model yourself, with support for screenshots, voice notes or screen recordings alongside code. However, activating only some of the parameters does not eliminate the high memory requirements: all experts remain in memory during operation.

Release card

Muse Spark Open

Meta

open weights
Specifications
The model has approximately 816 billion parameters, of which approximately 42 billion are active for each token.
Context
The context window holds up to 1 million tokens across text, images, audio and video combined.
Inputs
The model accepts text, images, audio and video and produces text output.
Availability
The weights are available to download on Hugging Face in BF16 and NVFP4 formats. The license allows free use with the mandatory attribution “Built with Muse Spark Open”. Providers with annual revenue above 20 million dollars need a separate agreement with Meta to offer the model via an API.
According to the sources, it is suitable for
  • Reading and editing code across project files.
  • Debugging code during long tasks involving tool calls.
  • Working with code supplemented by screenshots, voice notes or screen recordings.
Documented limits
  • Deployments with limited memory capacity: all experts remain in memory, and the BF16 weights take up approximately 1.5 TB.
  • The NVFP4 variant is intended for NVIDIA Blackwell GPUs.

According to Meta, the model outperforms Qwen 3.8, GLM 5.3 and Kimi K3 in several categories of the cited benchmarks, but significantly trails the paid Muse Wonder Full models.

The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →

What was added since the original report

Verified updates

  1. New verified information

    The weights are available on Hugging Face.; According to Meta, the model runs on two Mac Studio computers with 512 GB of unified memory each.; The license requires the attribution “Built with Muse Spark Open”.; API providers with annual revenue above 20 million dollars need a separate agreement.

    • The weights are available on Hugging Face.
    • According to Meta, the model runs on two Mac Studio computers with 512 GB of unified memory each.
    • The license requires the attribution “Built with Muse Spark Open”.
    • API providers with annual revenue above 20 million dollars need a separate agreement.

Two audiences, two different impacts

What this means

01

For individuals

A developer can combine code with images or audio when submitting programming tasks. However, running the model locally requires substantially more hardware than a typical personal computer; Meta describes a configuration of two Mac Studio computers with a total of 1 TB of unified memory.

What to do Before downloading the weights, compare the available hardware with the configurations described and allow for approximately 1.5 TB for the BF16 variant.
More practical updates →
02

For a business

A company planning to offer the model via an API must account for the licensing terms: mandatory attribution and, for annual revenue above 20 million dollars, a separate agreement with Meta.

Risks and compliance
What to decide Before integrating the model into a service via an API, check the required attribution and whether the company is subject to the separate agreement requirement.
More business impacts →
BF16 Hugging Face M5 Ultra Mac Studio Meta MoE Muse Spark Open NVFP4

Check the original

Event sources

clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.

2
Hugging Face Blog primary source · first detected Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta The Decoder (daily AI news) independent context Muse Spark Open is Meta's answer to Qwen, GLM, and Kimi in the open-weight race