NVIDIA Nemotron 3.5 Super VL
NVIDIA’s open-weight vision-language model supports coding, visual reasoning, document understanding and agents that use tools.
Muse Spark Open
dev.meta.ai
Meta’s open-weight multimodal model supports coding, tool workflows and long-context applications through self-hosting or hosted inference.
Visit website
Pricing
Free
Available on
Self-hosted, Hosted API
Category
Llm Models
What is Muse Spark Open?
Muse Spark Open is Meta’s downloadable model for developers building coding assistants, multimodal applications and agents. It accepts text, images, audio and video and generates text. Its sparse mixture-of-experts architecture contains 816 billion total parameters, with 42 billion active per token, and supports a 1,048,576-token context window.
Teams can serve the model on their own infrastructure or call a hosted provider without provisioning GPUs. Meta’s deployment documentation covers providers including Together AI, Fireworks AI and Baseten. This gives developers a choice between managing inference infrastructure and integrating a provider endpoint.
Key features
Multimodal input
Process text, images, audio and video within an application that returns text responses.
Long-context processing
Use a context window of 1,048,576 tokens for applications working with extensive supplied material and conversation context.
Coding and tool workflows
Supports coding, sequential tool calls and tool-failure recovery when connected to application tools.
Reasoning-effort control
Choose low, medium, high or xhigh; xhigh is the default. Other values return an error.
Downloadable checkpoints
Choose BF16 or NVFP4 weights to match the deployment. After downloading, local inference can run offline without an inference API key.
Hosted inference options
Integrate provider endpoints using the appropriate model identifier, API key and request format instead of managing a local serving cluster.
Who is it for?
Developers building coding assistants
Connect the model to repository tools to suggest changes and help investigate software issues.
Teams interpreting visual material
Build an application that answers questions about screenshots, charts and document images.
Infrastructure teams evaluating private inference
Compare self-hosting requirements with a hosted endpoint for an internal application.
How to use Muse Spark Open
- Review the Muse Model License and Usage Policy, then choose self-hosting or a hosted provider.
- For hosted access, select a documented provider and configure its API key and model identifier.
- For self-hosting, log into Hugging Face, review access conditions and submit the required identity and organization form before attempting a download.
- Once repository access is granted, prepare Python 3.10 or later, the Hugging Face CLI, sufficient storage and a supported runtime. Download a suitable checkpoint and check its tokenizer and chat template.
- Start the selected runtime or connect to the provider endpoint, send a small request and choose a supported reasoning-effort setting.
Muse Spark Open pricing
Free
Weights carry no license fee under the Muse Model License and Usage Policy. Self-hosting infrastructure and hosted inference have separate costs. Together AI lists paid token-based access; consult its pricing page for current service terms.
Check current pricing ↗What to know before you use it
- Hugging Face downloads require login, access-condition review and an identity/organization form; access is not guaranteed. Self-hosting is substantial: BF16 weights occupy approximately 1.65 TB and NVFP4 approximately 537 GB, excluding cache, activations and runtime overhead. Documented recipes use eight and four B300 GPUs respectively.
- Attribution and Usage Policy compliance are required. Model-as-a-Service providers must notify Meta and obtain a separate agreement if aggregate licensee-and-affiliate gross revenue exceeds USD 20 million in any consecutive 12 months. This restriction excludes internal use and products or services that are not Model as a Service.
- Hosted integrations expose different capabilities. For example, the documented Fireworks endpoint accepts text and images; confirm endpoint support before relying on audio or video input.
- Multi-step reasoning can fail, and language quality may degrade outside the strongly supported set.
Frequently asked questions
Can Muse Spark Open run offline?
Yes. Once the weights are downloaded, local inference can run offline without an inference API key. Initial repository access and downloading are separate prerequisites.
Do 42 billion active parameters mean only those weights need loading?
No. The active count describes per-token computation. Serving still requires loading the full model weights.
Can I use it without provisioning GPUs?
Yes. Meta documents hosted provider APIs, including Together AI, Fireworks AI and Baseten, with provider-specific authentication and model identifiers.