Meta has announced two models: Muse Code and Muse Spark 1.2, designed to improve code generation, support long-sequence agentic tool calling, and handle end-to-end developer workflows.
Focus of Spark 1.2
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1. According to Meta, the release includes improvements in code generation quality, complex debugging, repository understanding, and support for end-to-end developer workflows. The company says it significantly scaled up training compute for coding tasks and expanded the diversity of training environments.
The model was also trained extensively on long-horizon coding tasks such as whole-repository generation, large end-to-end projects, and auto-research workflows.
Co-training with Muse Code
Muse Spark 1.2 was co-trained with Muse Code so the pair performs optimally when used together for coding tasks. Training included rejection-sampled harness trajectories and recipe optimizations targeting goals, compaction, and subagents, and the Muse Code toolset was integrated to maximize harness compatibility.
Example output and prior version
Meta published an example SVG produced by Muse Spark 1.2: a pelican riding a bicycle. They note that the Muse Spark 1.1 pelican was shown on July 9, 2024, and judge the 1.2 pelican to be a small but material improvement.
Pricing and model IDs
Meta is offering Spark 1.2 under two different model IDs with distinct pricing:
- muse-spark-1.2: $1.25 per million input tokens and $4.25 per million output tokens.
- muse-spark-1.2-contributor: $0.10 per million input tokens and $0.20 per million output tokens; this discounted tier requires that users allow Meta to use their data “to improve our products.”
The announcement situates these prices relative to other models, noting the contributor tier is a substantial discount compared with standard commercial rates.
Why this matters
The release underscores a broader industry focus on handling long-sequence, agentic tool calling and on delivering models that can manage large, multi-step coding workflows. By co-training Muse Spark 1.2 with Muse Code and expanding compute and training environment diversity, Meta aims to make these models more reliable and useful for sizeable, complex coding tasks.
Tags (from Meta's post)
ai, generative-ai, llms, meta, llm-pricing, pelican-riding-a-bicycle, llm-release, coding-agents
