What is Muse Spark 1.1?
Muse Spark 1.1 is a hosted reasoning model from Meta, available through the Meta Model API. It is designed for applications that need more than a single text response: an application can provide multimodal inputs, ask the model to reason through a complex task, call external tools, use search grounding, and coordinate several steps in an agentic workflow.
The model accepts text, images, video, audio, and PDFs, but its native response format is text. That distinction matters: Muse Spark 1.1 can analyze media, yet it is not an image, video, audio, music, or speech generator. Its role is closer to a multimodal decision-making and orchestration engine than to a creative media-generation model.
Meta released Muse Spark 1.1 on July 9, 2026. The model remains listed as accessible on the Meta Model API Standard tier in public preview for US developers, although the supplied documentation identifies Muse Spark 1.2 and Muse Spark 1.3 as newer versions. Meta recommends Muse Spark 1.3 for new work, so Muse Spark 1.1 is most relevant when compatibility with the original version, existing testing, or a particular deployment requirement matters.
Where it fits in Meta’s lineup
Muse Spark 1.1 belongs to Meta’s Muse Spark family and is separate from the broader Llama model family. The model is offered as a managed API model rather than a self-hosted package. Its documented positioning emphasizes reasoning, multimodal understanding, software engineering, computer-use workflows, tool orchestration, and web-grounded applications.
For new projects, the newer Muse Spark versions may be the more appropriate starting point because Meta recommends Muse Spark 1.3. However, the available research does not provide a detailed feature-by-feature comparison between 1.1, 1.2, and 1.3. It would therefore be inaccurate to claim a specific benchmark, price, context-window, or capability advantage for the newer models without separate documentation.
Core capabilities and supported inputs
Muse Spark 1.1 supports the following input types:
- Text
- Images
- Video
- Audio
- PDFs and other supported document inputs
Its output is text. The model’s multimodal input makes it suitable for tasks such as explaining an image, reviewing a document, examining a video, interpreting an audio recording, or combining evidence from several media types. The output limitation means that an application needing a generated image, synthesized voice recording, music track, or video should use a separate specialized model or service.
The model also supports structured output. Structured output is useful when an application needs the response to follow a defined schema, such as returning extracted fields from a PDF, classifying a set of images, or producing a machine-readable plan for a software agent. The research confirms structured-output support, but it does not establish that Meta exposes a separate capability called JSON mode. Those should not be treated as interchangeable features without checking the API documentation for the exact request format.
Context window and reasoning
The verified context window is 1,048,576 tokens, or approximately one million tokens. A context window is the amount of information the model can consider within a request and its surrounding conversation or tool workflow. In practical terms, this makes Muse Spark 1.1 suitable for unusually large collections of source material, long codebases, extended transcripts, large document sets, and multi-step agent sessions.
A large context window does not automatically mean that every long prompt will be handled equally well. Applications still need to organize source material, identify the most relevant evidence, and manage the cost of processing large inputs. Long requests may also be more expensive and can make application-level testing important.
Muse Spark 1.1 supports reasoning-effort controls. These controls allow an application to choose how much reasoning work the model should devote to a request, where supported by the API. Higher reasoning effort may be useful for difficult coding, planning, or tool-selection tasks, while lower effort can reduce latency and output-token consumption for simpler operations.
The provider’s exact maximum output-token limit was not verified in the supplied documentation. Reasoning tokens count as output tokens for billing and consume the output-token budget. Developers should therefore confirm the current API limit and leave enough output capacity for both hidden reasoning work and the visible answer.
Coding, tools, and computer-use workflows
Tool and function calling are central to Muse Spark 1.1’s intended use. A tool call allows the model to request an operation from the surrounding application, such as querying a database, calling a business API, retrieving a file, running a search, or executing an action in a controlled environment. The application, rather than the model alone, normally performs the operation and returns the result to the model.
The model supports computer-use workflows through tools. This can help an agent interpret a task, decide which computer interaction is needed, and coordinate actions exposed by the host application. Computer use should not be understood as unrestricted control of a user’s machine. The actual permissions, available actions, confirmation steps, and safety boundaries depend on the tools that developers provide.
Muse Spark 1.1 is also aimed at software engineering. Suitable tasks include code generation, code explanation, debugging assistance, repository-level analysis, planning implementation steps, and coordinating development tools. Its million-token context can be useful when a task requires examining many files or a lengthy technical specification. However, generated code still requires testing, review, dependency checks, and security evaluation; the research does not provide a verified benchmark proving a particular level of coding accuracy.
Streaming is supported, allowing an application to receive output progressively instead of waiting for the complete response. This can improve perceived responsiveness, especially for long reasoning or tool-driven interactions. Prompt caching is also supported. Reusing cached prompt material can reduce the cost of repeated requests containing the same large instructions or reference content, although developers should confirm the provider’s current cache rules and eligibility requirements.
Search grounding and current information
Muse Spark 1.1 supports search grounding. In a grounded workflow, the model can use search results supplied through the provider’s search capability to answer questions with more current information than its underlying training data alone can provide. This is useful for research assistants, current-events queries, product information, and applications that need citations or external evidence.
Search grounding does not change the model’s knowledge cutoff. Meta’s documentation does not specify an exact training-data or knowledge-cutoff date for Muse Spark 1.1. Search results can provide current information during a request, but they do not establish that the model itself has learned that information permanently.
Web search grounding carries an additional documented cost of $2.50 per 1,000 search queries. Search results should still be checked for relevance, source quality, and completeness. Grounding can improve freshness, but it does not guarantee that every generated conclusion is correct.
Pricing and cost trade-offs
On the Meta Model API Standard tier, Muse Spark 1.1 is priced at:
| Usage type | Price |
|---|---|
| Input tokens | $1.25 per 1 million tokens |
| Cached input tokens | $0.15 per 1 million tokens |
| Output tokens | $4.25 per 1 million tokens |
| Web search grounding | $2.50 per 1,000 search queries |
These are usage prices rather than a fixed consumer subscription. Reasoning tokens are counted as output tokens, so a request that produces a short visible answer can still consume a larger billable output amount when the model performs substantial internal reasoning. Large multimodal inputs, long contexts, repeated tool calls, and search requests can also increase total usage.
Prompt caching materially lowers the listed input rate for eligible cached material, from $1.25 to $0.15 per million tokens. This makes repeated workflows with stable instructions or reference documents a better fit than one-off requests with constantly changing context. The model’s supplied editorial ratings describe its cost as favorable and its speed as strong, but those ratings are evaluations rather than Meta-published benchmark results.
Main strengths and limitations
The strongest practical case for Muse Spark 1.1 is the combination of a very large context window, multimodal input, reasoning controls, tool use, coding support, and search grounding. Few ordinary chat-style workflows need all of these capabilities at once, but complex agents often do. For example, an agent could inspect a long technical repository, read an attached design document, analyze a diagram, search for current documentation, and call project-management tools before returning a structured implementation plan.
Its limitations are equally important:
- It returns text rather than directly generating images, video, audio, music, or speech.
- The maximum output-token limit was not verified in the supplied research.
- The exact knowledge cutoff is not published in the supplied documentation.
- Fine-tuning and batch API support were not verified for this specific model.
- It is not documented as a self-hosted model.
- Search grounding adds a separate charge and still requires source evaluation.
- It is in public preview for US developers, so availability and behavior may change.
- It has been superseded in the Muse Spark family by versions 1.2 and 1.3, with Meta recommending 1.3 for new work.
When to choose Muse Spark 1.1
Choose Muse Spark 1.1 when an application needs a managed model that can combine multimodal understanding with long-context reasoning and external actions. It is particularly suitable for coding agents, document and media analysis, computer-use systems, research assistants with search grounding, and workflows that repeatedly process the same large reference material.
The model is also a reasonable option when input flexibility matters more than generated media. A team could use it to interpret an image or video and then return a textual report, or to inspect a PDF and produce structured fields for another system.
Another model type may be more appropriate when the main requirement is direct media generation, speech synthesis, transcription, embeddings, or self-hosted deployment. A smaller or less reasoning-intensive model may also be preferable for simple classification, short extraction tasks, or high-volume requests where latency and cost matter more than complex planning. For a new Meta project, the newer Muse Spark 1.3 should be evaluated first because Meta recommends it for new work, while retaining 1.1 may make sense for compatibility or version-specific testing.
Bottom line
Muse Spark 1.1 is best understood as a text-output reasoning engine for multimodal, tool-using applications. Its defining technical advantages are the one-million-token context window, support for text, images, video, audio, and PDFs, structured output, reasoning controls, coding assistance, computer-use workflows, streaming, caching, and optional search grounding.
It is not an all-purpose media generator, and its preview status and superseded position should be considered before adoption. For teams that need long-context analysis and agentic orchestration through Meta’s API, it remains a capable option. For new deployments, however, the newer Muse Spark versions deserve priority evaluation, and any production decision should verify current availability, output limits, pricing, and feature support directly in Meta’s documentation.