Meta released the weights for Muse Spark Open on October 15, 2026, giving businesses building custom AI assistants a new way to run and adapt the model on their own infrastructure. Meta is targeting coding and multi-step work, with hosted access also available through partners.
Weights are downloadable model files containing the settings learned during training. Having them lets a team choose where the model runs and pursue further training, subject to Meta’s licence. The Hugging Face download page requires users to sign in and review access conditions.
This is a separate release from Muse Wonder, which RohitAI covered earlier this week.
Running it yourself takes substantial hardware
Meta’s download guide puts the reduced-precision NVFP4 version at about 537 GB of weights, before the extra memory needed to handle requests. One documented setup uses four NVIDIA B300 GPUs. The full BF16 version contains about 1.65 TB of weights.
Although only 42 billion of the model’s 816 billion parameters are active for each token—a unit of input or output—the remaining weights still have to be stored. The smaller active count does not make this a small download.
The practical benefit is control over deployment, not established cost savings. A team running the model itself still needs to budget for computing capacity and operating the service.
Hosted access has different limits
Meta documents text, image, audio and video inputs, with text output. Fireworks’ hosted service lists text and image inputs, so the model’s full range of inputs should not be assumed available through every provider.
Fireworks lists prices of $1.25 per million input tokens, $0.15 for eligible cached input and $4.25 for output. Its serverless offer is a limited preview through October 29, 2026, running on US infrastructure, with the usual US-capacity surcharge waived.
For a business without suitable hardware, hosted evaluation avoids an upfront capacity purchase. But anyone planning to use Fireworks beyond that preview should establish the next hosting arrangement and its costs; the preview rate card is not a dedicated-capacity quote.
What commercial users need to check
The custom Muse Model License permits use and adaptation without a licence fee under its stated conditions. Users must follow the usage policy, preserve required notices and display “Built with Muse Spark Open” in a permitted location, such as documentation or a website.
Its separate-agreement requirement applies to qualifying model-as-a-service businesses—those providing others with access to models through an API or similar interface—when combined gross revenue of the licensee and its affiliates exceeds $20 million in any consecutive 12 months. They must notify Meta and reach a separate written agreement before offering the model or derivatives through that business.
Section 3 expressly excludes internal use, research and commercial activities, including product features, that are not model-as-a-service. The threshold is not a blanket fee for every company above $20 million.
Reporting note: This AI-assisted report draws on Meta’s release, documentation and licence, plus Fireworks’ public listing, checked on October 15, 2026. RohitAI has not tested the model or hosted service.