Muse Spark Open: is Meta’s new AI free, and can you run it locally?

What is Muse Spark Open, and is it the same as Muse Wonder?

Muse Spark Open is the open-weight release of Muse Spark, aimed at coding and agent workflows. Meta lists a context window of roughly one million tokens. It is a separate release from Muse Wonder and Muse Wonder Full, announced on October 12; do not treat the names as interchangeable.

Meta distributes the weights under the Muse Model License with limited commercial terms. It lists downloads through Hugging Face and cloud catalogs, plus hosted access from partners including OpenRouter, Together AI and Fireworks AI. Check the selected provider’s current access conditions.

Sources: Meta: Introducing Muse Spark Open · Meta: Muse Spark Open developer documentation · Meta: Introducing Muse Wonder and Muse Wonder Full

Is Muse Spark Open free? What does running locally require?

A downloadable model is not a promise of a free hosted chat. Meta’s desktop example uses two M5 Ultra Mac Studios. Its documentation lists approximately 537 GB of weights for the NVFP4 checkpoint, before additional serving memory. This is substantial infrastructure, not a normal laptop installation.

Using a hosted service avoids buying that hardware. However, the service can set its own pricing, access limits and supported features. For example, Groq’s launch documentation describes limited preview access, text and image inputs, and a 131,072-token context limit. Model specifications and a particular service’s offering can differ.

Sources: Meta: Introducing Muse Spark Open · Meta: Muse Spark Open developer documentation · Groq: Muse Spark Open availability and limits

Can Muse Spark Open analyze images and documents?

The model card lists interpreting images, charts and documents among its uses. Its output is text, so understanding a picture should not be confused with generating a new picture. Direct PDF uploads and file handling depend on the application around the model; check those features before choosing a service.

A useful first task is a two-source comparison. Take a fictional product sheet and a photograph of the product. Ask the assistant to separate what the sheet states from what the photograph actually shows. A label on paper is evidence for a written claim, not proof that a physical item has that property.

Sources: Meta: Muse Spark Open model card

A practical prompt for checking a document against an image

Prepare a short sample you understand: a fictional backpack sheet that says “20 litres, two side pockets, available in blue,” plus a photograph you have permission to use. If your chosen interface accepts images, attach it; otherwise use the text exercise alone and do not ask for visual judgments.

This exercise is designed to reveal unsupported assumptions. A photograph may show pockets, but it usually cannot verify internal capacity. The assistant should preserve that distinction instead of turning every printed claim into an observed fact.

• Accuracy: does each detail match the supplied material?
• Evidence: does the answer distinguish visible details from written claims?
• Restraint: does it say what it cannot verify rather than inventing an answer?

Compare the supplied product text and photograph. Make three lists: claims stated in the text, details clearly visible in the image, and claims you cannot verify. For each answer, identify its source. Do not infer dimensions, capacity or materials from appearance. Then write a 60-word product description using only supported details, followed by two questions for the seller. If no image is available, say so and analyze only the text.

How to choose an AI chat service for this kind of work

Start with a small repeatable job, not the biggest model name. If you compare product information every week, save a sample brief and a checklist of mistakes that matter to you. Try the same material in the services you can access and compare the corrections required.

Before paying, check whether the interface accepts your file type, whether its answer can point back to the source, and what happens when the source is unreadable. A long context window alone cannot tell you whether tiny text, a confusing table or a photographed label will be interpreted correctly.

• Occasional questions: a ready-to-use chat interface may be enough.
• Repeated team workflows: compare cost per completed task and review time.
• Self-hosting: assess hardware, maintenance and the license before planning a deployment.

A simple checking routine before relying on the answer

Keep the original material beside the answer. Underline every factual statement in the draft, then trace it back to a specific line or visible feature. Remove statements whose support you cannot find. Check numbers, units, names and exclusions separately: these often matter more than polished wording.

Finally, change one detail in the sample and repeat the task. For instance, change the written capacity but leave the photograph unchanged. A useful result updates the written claim while keeping the visual evidence separate. This is a suggested checking method, not a published performance score for Muse Spark Open.