Muse Spark 1.2 for Coding

Muse Spark 1.2 is Meta's first model tuned explicitly for long-horizon, multi-agentic coding workflows. It ships behind the Muse Code CLI released 2026-08-05.

This page covers what Muse Spark 1.2 does well on code tasks, where it fits alongside your existing tools, and the routing decisions worth making.

Benchmark scores are not published at launch — verify at https://developer.meta.com/ai/products/muse-code/.

Where Muse Spark 1.2 Is Strong on Code

Muse Spark 1.2 is designed for the tasks a single-shot model handles poorly: multi-file refactors, long test-loop debugging, and end-to-end feature slices.

The multi-agent-by-default architecture assigns worker agents to sub-tasks and a reviewer agent to critique the diff before it lands. That is the same pattern a good human PR review follows.

Persistent memory means you can close a laptop mid-refactor and pick up where you left off — a real gap in single-session CLIs.

• Multi-file refactors and long-horizon feature slices
• Parallel workers + background reviewer
• Persistent memory across sessions
• Event log for every edit and decision

Task Routing: When to Pick Muse Spark 1.2

Route long-horizon tasks to Muse Spark 1.2 through Muse Code. That includes refactors, multi-file features, and end-to-end bug fixes.

Route quick edits and single-file completions to your IDE-embedded assistant. The overhead of a multi-agent loop is not worth it for one-liners.

Route sensitive workloads to Standard tier only. Contributor is fine for OSS and prototypes but not for client code.

• Muse Code + Muse Spark 1.2: long-horizon work
• IDE assistant: single-file, quick edits
• Standard tier: sensitive workloads
• Contributor tier: OSS + prototypes

Pair with Muse Code for Full Value

Muse Spark 1.2 through the raw API is functional. Muse Spark 1.2 through Muse Code is where the multi-agent workflow, event log, and persistent memory actually live.

For a team eval, use Muse Code. That is the intended surface and the one Meta will keep tuned as the model evolves.

In our engagement with client engineering teams, we treat model + agent as one integration decision — not two — because tuning happens at that boundary.

• Muse Code is the intended surface
• Event log + memory live at the agent layer
• Evaluate model + agent as one integration
• Tuning happens at the boundary

Known Gaps on Code Tasks

SWE-bench and HumanEval scores are not published — verify at https://developer.meta.com/ai/products/muse-code/. Do not assume parity with Claude Fable 5 or GPT-5.6 Sol on any specific eval.

The full built-in tools list beyond "file edit, tool call" is not documented. Whether Muse Code supports MCP, shell exec, or web fetch out of the box is unpublished.

For language coverage — languages beyond mainstream Python / TypeScript / Go / Rust may see uneven quality. Test your primary language stack before committing.

• No published SWE-bench or HumanEval scores
• Full built-in tools list not documented
• Language coverage: test your stack first
• Do not assume parity with rivals