Meta has introduced Muse Spark 1.2, an enhanced version of its coding-focused model Muse Spark 1.1. This update focuses on improvements in code generation, debugging, codebase understanding, and overall developer workflows. Significant scaling of training compute for coding tasks, along with diversity in the training environment, has been implemented. Muse Spark 1.2 retains strength in areas beyond coding, particularly in general agent tasks.
To optimize performance, Muse Spark 1.2 was co-trained with Muse Code, ensuring enhanced coding usability when utilized in tandem. The training process employed sampling and optimization techniques aimed at improving goal-directed behaviors, code compaction, and agent interactions. The model's training extensively covered long-horizon coding tasks, which include generating entire repositories, executing large-scale end-to-end projects, and facilitating automated research tasks. The overall performance improvement is suggested to be tangible, though described as incremental.
In terms of pricing, the model is offered under two different IDs: muse-spark-1.2 at $1.25 per million inputs and $4.25 per million outputs, which aligns closely with competitors like Gemini 3.6 Flash. Alternatively, users can opt for muse-spark-1.2-contributor at a significantly reduced rate of $0.10 per million inputs and $0.20 per million outputs, provided they consent to allow Meta to utilize their data for product enhancements. This discount positions the model competitively next to GPT-5.6 Luna and Gemini 3.1 Flash-Lite options.