Muse Spark 1.2 is Meta's coding-focused flagship model, released August 5, 2026 alongside Muse Code, a beta terminal coding agent. It significantly scales training compute on coding tasks and expands training environment diversity, improving code generation, complex debugging, codebase understanding, and end-to-end developer workflows. With a 1M context window, higher first-attempt accuracy, and more reliable tool calling, it is available today in Muse Code and through the Meta Model API with expanded global access.

Core Features

Coding-focused training: Significantly scaled compute on coding tasks with broader environment diversity for code generation, debugging, and codebase understanding.
1M context: Long-running tasks run start to finish in one session without context splitting.
Higher first-attempt accuracy: Optimized for real coding workflows with fewer failed attempts.
Reliable tool calling: Built for agentic use with consistent function and tool invocation.
Co-trained with Muse Code: Trained alongside the terminal agent using rejection-sampled harness trajectories, goals, compaction, and subagents for best harness compatibility.
Dual tiers: muse-spark-1.2-contributor (used to improve Meta products) and muse-spark-1.2 (not used for product improvement).

Model Specifications

Specification	Muse Spark 1.2
Focus	Coding, debugging, codebase understanding, developer workflows
Context	1M tokens
Access	Muse Code (beta), Meta Model API
Tiers	Contributor and non-contributor
Open weights	Announced for coming weeks (per Meta)

Use Cases

Terminal coding agents: Muse Code uses Muse Spark 1.2 to build, debug, and ship code with multiple coordinating agents.
Complex debugging: End-to-end developer workflows on large codebases.
General agents: Maintains strength in general agent tasks alongside coding.
Autonomous GitHub bots: Build GitHub agents on the Meta Model API with Muse Spark and OpenCode.

Advantages

Meta frontier coding: The model behind Muse Code, Meta's first terminal coding agent, benchmarked close to Claude Opus 5.
1M context: Whole-repo, start-to-finish sessions in one call.
Harness-native: Co-trained with Muse Code's toolset for the best agent experience.

Tips

Pair with Muse Code for the intended harness integration; the model was co-trained with it.
Choose the non-contributor tier if you do not want your prompts used to improve Meta products.
Use the 1M context for whole-repo tasks to avoid manual context splitting.

Conclusion

Muse Spark 1.2 is Meta's answer to the coding-agent frontier: a 1M-context, coding-focused model co-trained with its own terminal agent. If you want Meta's frontier coding capability through a first-party harness or API, it is the flagship choice.

Related: Claude 3.5 Sonnet and Claude 3 Haiku. Hub: models.