Muse Spark 1.3 is Meta’s model for agentic tasks, long-running coding work, and multimodal interaction. It is designed to retain context and track earlier results across extended workflows, handle messy or conflicting instructions, and request clarification when needed. Meta also highlights higher first-try accuracy and dependable tool use. The model can interpret images, video, and documents; its visual reasoning operates through an execution environment, allowing it to act on screenshots or clips rather than follow only scripted steps. The intended uses range from coding assistants to agents that browse, operate computers, or complete business workflows.
Benchmark results show gains over Muse Spark 1.2 in every listed comparison, including long-context retrieval, coding, and agentic tasks. On the displayed measures, version 1.3 scores 98.5 and 98.1 on two long-context tests, versus 66.3 and 55.5 for version 1.2, and 75.4 on the DeepSWE coding benchmark versus 55. It also scores 59.4 on codebase question answering and 88.8 on terminal coding. Results against other frontier models are mixed: Muse Spark 1.3 leads on some measures, while Opus 5 scores higher on several others. The model offers a one-million-token context window; listed pricing is $1.25 per million input tokens and $4.25 per million output tokens, with a cheaper contributor version whose prompts may be used to improve Meta’s products. Integration options include Meta’s API, OpenRouter, Google Cloud, and coding tools.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.