What Is Muse Spark?

What changed in Muse Spark 1.2

Meta describes 1.2 as a coding-workload update: significantly more training compute on code generation and debugging, more diverse training environments, and extensive training on long-horizon tasks up to whole-repository generation. The distinctive design choice is that the model was co-trained with the Muse Code harness from inception, so the agent and the model were tuned together rather than the agent being a wrapper over a general model.

Meta reports the model still generalizes to other coding agents, native tools, MCP servers, and custom skills. In one published stress test it sustained an iterative GPU kernel optimization across more than 1,000 tool calls.

• Context window: 1,048,576 tokens (1M); maximum output 131,072 tokens
• Reasoning effort levels from none to ultra
• Closed weights; open-sourcing has been announced as intent but nothing has shipped
• Served via the Meta Model API with OpenAI-SDK-compatible endpoints

The two pricing tiers, and why they matter

The standard muse-spark-1.2 tier is priced at $1.25 per million input tokens, $4.25 per million output tokens, and $0.15 per million cached input tokens, with no training on your data and zero-data-retention available on request.

The muse-spark-1.2-contributor tier is dramatically cheaper, around $0.10 input and $0.20 output, but it is rate-limited on a rolling window and, critically, it grants Meta training rights on the data you send. For a coding agent that may be reading tool responses from your email, documents, or internal systems over MCP, the tier you select decides whether that content can end up in a training corpus. This is a documented, tier-specific policy rather than a blanket claim about Meta; always confirm the current terms for your tier.

Muse Spark vs Muse Code

The names travel together but they are different layers. Muse Spark is the model: it generates the plans, code, and tool calls. Muse Code is the agent harness: the terminal program that manages sessions, sub-agents, worktrees, approvals, and MCP connections, and sends requests to Muse Spark. You can call Muse Spark directly through the Meta Model API without ever installing Muse Code, and Muse Code is the primary consumer of Muse Spark 1.2.

What Muse Spark's context window means for data exposure

A 1M-token context window is the headline capability, and it cuts both ways. It lets the agent hold a large codebase in one session. It also means a single session can accumulate an enormous amount of whatever the agent has been given access to, including tool responses from connected business systems. The wider the window, the more valuable minimization becomes: sending the model three redacted fields instead of a full mailbox thread keeps both the exposure and the noise down.