What is Muse Spark 1.2?
Muse Spark 1.2 is a coding-focused reasoning model from Meta. It is intended for software-engineering tasks that extend beyond generating a short function or answering a programming question. Typical workloads include inspecting a large repository, planning changes across multiple files, debugging an unfamiliar codebase, running tests through tools, and iterating until the requested change is complete.
The canonical model identifier is muse-spark-1.2. Meta also lists muse-spark-1.2-contributor, a separate discounted tier with different data-use terms. The contributor identifier should therefore be treated as a distinct commercial option rather than simply another name for the standard model.
Muse Spark 1.2 is still listed and accessible through Meta Model API and Muse Code, and it is also available through OpenRouter according to the supplied model information. Meta’s current catalog positions Muse Spark 1.3 as the latest and recommended version, making Muse Spark 1.2 a previous-generation option rather than the default choice for new deployments.
Why it is designed for coding agents
A coding agent is an application that combines a language model with tools such as file search, command execution, test runners, issue trackers, or web search. Instead of receiving one prompt and returning one answer, the model can make a plan, inspect information, request an operation, review the result, and continue with another step.
Muse Spark 1.2 is designed around this longer loop. Meta describes support for whole-repository generation, planning, goal conditioning, context compaction, asynchronous tool calls, and parallel tool calls. In practical terms, an agent can divide work such as repository inspection, dependency analysis, and test discovery into separate operations, then use the returned results to guide the next stage.
This does not mean the model independently has unrestricted access to a computer or codebase. The surrounding application must provide the tools and decide which actions are permitted. The model’s documented role is to reason about the task, produce text and structured responses, and request tool calls that the application can execute.
A 1,048,576-token context window
Muse Spark 1.2 has a documented context window of 1,048,576 tokens. A token is a unit of text processed by the model; the exact number of words represented by a token varies by language and content. A context window is the amount of input and conversation state the model can consider in a request.
This capacity is particularly relevant to repository-scale work. An application may be able to include substantial source code, documentation, configuration files, dependency information, previous tool results, and task history without immediately splitting the work into many unrelated sessions. The model is therefore a better fit for long-running development workflows than a model intended primarily for short prompts.
The large context window is not a guarantee that every large repository should be sent in full. Applications still need to select relevant files, manage token costs, protect sensitive information, and control how tool results accumulate. The supplied documentation does not specify an exact maximum output-token limit or an exact knowledge-cutoff date for Muse Spark 1.2.
Supported input and output modalities
Muse Spark 1.2 accepts text, images, video, audio, and PDF inputs. This gives coding and research applications more options than a text-only coding workflow. For example, a developer could provide a screenshot of a user-interface bug, a chart from a monitoring system, a PDF specification, or an audio-visual explanation alongside written instructions.
The model returns text output. It does not generate native images, video, speech, music, transcriptions, or embeddings. Its multimodal capability is therefore about understanding supplied media and using that information in reasoning or tool-driven tasks, not about producing media files.
Meta’s research materials describe multimodal reasoning, visual coding, audio-visual understanding, and tool-assisted perception. These are provider-described capabilities; the supplied research does not provide independent benchmark results establishing how Muse Spark 1.2 compares with other multimodal models on those tasks.
Reasoning, tools, and structured responses
The model is intended for tasks requiring planning across multiple steps. Its reasoning-oriented design is useful when the correct answer depends on inspecting evidence, keeping track of constraints, and revising a plan after a tool returns new information. Repository debugging is a representative example: the agent may inspect an error, trace related code, identify a likely cause, apply a change, and run tests.
Muse Spark 1.2 supports tool calling, including parallel tool workflows, as well as streaming. Streaming allows an application to receive output progressively instead of waiting for the entire response. Prompt caching is also supported, which can reduce the cost of repeatedly supplying unchanged context in supported workflows.
The model supports structured output. Structured output means an application can request a response that follows a defined schema, such as a list of proposed file changes, a test plan, or a machine-readable task status. The reviewed documentation does not independently verify a separate legacy JSON-mode feature beyond structured output, so those capabilities should not be treated as interchangeable without checking the specific API implementation.
Web-search grounding is available through Meta Model API. It can provide current external information during a request, but it does not change the model’s underlying training knowledge cutoff. Since no exact cutoff date was verified, applications that require current facts should use web grounding or another explicit retrieval system and should validate important results.
Pricing and commercial tiers
The standard Muse Spark 1.2 price is $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens. Input tokens are the material sent to the model, while output tokens are the generated response. Cached input pricing applies when supported content can be reused according to the API’s caching rules.
Meta also lists a contributor tier using the identifier muse-spark-1.2-contributor. Its listed prices are $0.10 per million input tokens, $0.002 per million cached input tokens, and $0.20 per million output tokens. The contributor tier may use prompts and completions to improve Meta products, while the standard tier is described in the supplied research as not being used to improve products.
These prices are usage-based rather than a monthly end-user subscription. Actual project cost depends on the amount of repository context, tool output, conversation history, caching, and generated material sent through the service. A large context window can reduce the need to create separate sessions, but it can also increase input-token usage if an application repeatedly sends extensive context without effective caching or selection.
Main strengths and trade-offs
- Long-horizon software work: The million-token context and agent-oriented design suit repository-scale analysis, multi-file refactoring, and extended debugging sessions.
- Multimodal understanding: Text, image, video, audio, and PDF input can be combined with coding or research instructions.
- Tool-oriented workflows: Tool calling, parallel tool calls, streaming, structured output, and web-search grounding support integration into development agents.
- Flexible deployment choices: The model remains available through Meta Model API and Muse Code, with OpenRouter also listed as an access route.
- Cost distinction between tiers: The contributor tier is substantially cheaper, but its data-use terms differ from the standard tier.
The main trade-off is that Muse Spark 1.2 is no longer Meta’s recommended latest model. New projects may therefore prefer Muse Spark 1.3 if they want the current version, although the supplied research does not provide enough detail to quantify the newer model’s price, speed, or quality advantage.
The model also produces text only, so it is not suitable when the application needs native image, video, or audio generation. Its API documentation does not verify an exact maximum output size, fine-tuning availability, or batch API support. Those omissions may matter for production planning and should be checked against the current service documentation before implementation.
When to choose Muse Spark 1.2
Muse Spark 1.2 is a reasonable choice when the main problem is long-running, tool-driven software engineering and the application benefits from keeping substantial context in one workflow. Suitable examples include:
- Repository-scale debugging where the agent must examine many related files.
- Multi-file refactoring with dependency and test awareness.
- Migration planning across legacy code, configuration, and documentation.
- Coding assistants that inspect files, execute commands, run tests, and iterate.
- Development workflows that combine screenshots, PDFs, diagrams, or recordings with source code.
- Structured agent outputs that must be consumed by another application.
Another model type may be more appropriate for a short coding question where low latency and minimal cost matter more than a very large context. A dedicated image, video, speech, transcription, or embedding model is the better fit when native media or vector output is required. A current sibling such as Muse Spark 1.3 may be preferable for new work because Meta recommends it as the newer version, but the supplied sources do not establish a detailed apples-to-apples comparison.
Release status and known limitations
Muse Spark 1.2 was released on August 5, 2026. It remains available and has no verified shutdown or deprecation date in the supplied research. Its status is best described as available but previous-generation, with Muse Spark 1.3 recommended for new work.
Important limitations are the absence of a verified exact knowledge-cutoff date and maximum output-token limit, the lack of independently supplied benchmark results, and the fact that several operational details—such as fine-tuning and batch API support—were not verified. Media understanding should not be confused with media generation, and web-search grounding should not be confused with a permanently updated model.
Overall, Muse Spark 1.2 is most distinctive as a large-context coding model for agentic workflows. Its value comes from combining repository-scale context, multimodal input, reasoning, tool use, and structured text output. It is less compelling when the task is simple, latency-sensitive, media-generative, or better served by Meta’s newer recommended model.