Meta unveils Muse Spark 1.2 with Muse Code co-training and long-horizon coding benchmarks

Meta released Muse Spark 1.2, a coding-focused update to Muse Spark 1.1 that improves code generation, complex debugging, codebase understanding, and end-to-end developer workflows. The company significantly scaled training compute on coding tasks and expanded the diversity of training environments while maintaining strength in general agent capabilities.

Muse Spark 1.2 was co-trained with Muse Code, Meta’s new coding agent, to ensure the model performs best when paired with the agent. Training incorporated rejection sampled harness trajectories and recipe optimizations for goals, compaction, and subagents, alongside integration of the Muse Code toolset to maximize harness compatibility.

Muse Spark 1.2 was extensively trained on long-horizon coding tasks, including whole-repository generation, large end-to-end projects, and auto-research. Meta argues that long-sequence agentic tool calling is becoming a defining characteristic of modern models.

Muse Spark 1.2 is offered as two model IDs. The standard ID, muse-spark-1.2, is priced at $1.25 per million input tokens and $4.25 per million output tokens. A second ID, muse-spark-1.2-contributor, is priced at $0.10/$0.20 if users consent to allow Meta to use their data to improve products, representing a substantial discount relative to the standard pricing.