Understand a codebase
Connect relevant files, dependencies and the change you want to make.
Muse Spark
A model for understanding complex work and carrying it through.
Meet Spark.
Muse Spark is Meta's cloud-hosted AI model for software development, computer use, tool calling and long multi-step work. It powers Muse Code and much of the consumer Muse product. Developers can also use Meta Model API, an interface that lets their software send requests to the model.
What Spark is good at
Use software and tools
Work through the interfaces supplied by the surrounding application.
Stay with a longer task
Use context across multiple steps, then inspect the outcome.
Read beyond text
Reason about images and documents alongside your instructions.
Find and fix a broken checkout.
Follow a task from the repository to a reviewable correction.
- 1
Understand the failure
Connect the user’s report with the address form, relevant code and expected checkout behavior.
- 2
Form a hypothesis
Consider whether the empty optional field is being treated as a missing required value.
- 3
Choose the next check
Ask the development tools to reproduce both the empty-field and completed-field cases.
- 4
Evaluate the result
Use the tool output to decide whether to revise the hypothesis or propose a correction for review.
Illustrative reasoning process. Muse Code supplies the editing and testing workflow.
Where it fits in Muse
| Consideration | Spark | Glimmer |
|---|---|---|
| Runs | Hosted service | Your hardware |
| Typical fit | Complex reasoning and tool-based work | Local files and self-managed workflows |
| You provide | Context, tools and permissions | Hardware, runtime, tools and controls |
| Tradeoff | Data handling depends on the hosted service | Hardware needs and capability limits |
Real examples
1. Receive task2. Generate code3. Run application
Browser-Based Website Development
An AI workflow that creates a website, opens it in a browser, inspects the result and makes corrections.
View example
1. Receive task2. Generate game code3. Launch in browser
Iterative Game Development
Builds a game, runs it, playtests through a browser and improves based on observed failures.
View example
1. Define brief2. Plan work3. Build in parallel
Multi-Agent Product Studio
Product, frontend, backend and writing roles coordinate on a software project.
View exampleTechnical architecture
Spark runs through Meta Model API (and Muse Code). Applications send requests to Meta's hosted service, where the model can plan, call tools, read multimodal inputs and continue across long contexts.
Specifications and limitations
- Current version
- muse-spark-1.3
- API pricing — checked September 29, 2026
- Standard: $1.25 input / $4.25 output per million tokens. Contributor: $0.10 input / $0.20 output, with permission to train on requests. Cached input is discounted; web search is billed separately. Check current terms in the footer resources.
- Context window
- 1,048,576 tokens — small units used to measure model input and output
- Availability
- Meta Model API · Muse Code · Muse
- Base URL
- https://api.meta.ai/v1
Keep in mind
- Runs in the cloud — requests are processed by Meta. Standard API prompts and completions are not used for training; Contributor requests permit training. No-training is not the same as zero retention.
- Audio understanding in Spark 1.3 is not fully supported; use Spark 1.2 for audio understanding or Voice Transcribe for speech-to-text.
- Useful results still depend on good tooling, approvals and review
- Not a substitute for human judgment on consequential actions