Meta's Muse Spark 1.3 targets agentic coding with higher first-attempt accuracy

Meta has released Muse Spark 1.3, a model trained for long-horizon, agentic workflows and competitive coding. It tracks context and prior results, handles messy inputs, and asks for clarification when needed. With native multimodal perception, it can process video, images, and documents, and its visual reasoning runs through a real execution environment. Benchmarks show competitive performance with frontier models on coding evals. Pricing starts at $0.10 per million input tokens for the contributor version, with a 1M context window.