Meta Unveils Muse Spark 1.3 with 1 Million Token Context Support

Meta ($META) has unveiled its inference-based artificial intelligence (AI) model 'Muse Spark 1.3,' designed to improve coding and agent workflow efficiency. The new model supports a 1 million token context window while maintaining API pricing at previous levels, according to Meta's official announcement on February 2.
Muse Spark 1.3 will be available through Muse Code and Meta Model API, processing text, images, and video inputs while tracking previous context and results for extended workflows. Internal benchmarks show Muse Spark 1.3 reduces tool calls by approximately 20% and token usage by about 25% in coding tasks compared to its predecessor Muse Spark 1.2.
The 1 million token context allows the model to reference long documents, codebases, and conversation flows in a single input. Developers can include multiple files and work histories for instructions, while agent-based services can maintain longer instruction and result retention during multi-step tasks.
Artificial Analysis ranked Muse Spark 1.3 sixth among 636 intelligence models, with long context search scores of 98.5 and 98.1 in the same evaluation. The API pricing for Muse Spark 1.3 xhigh remains consistent with previous versions: $1.25 (approx. 1688 KRW) per 1 million input tokens, $4.25 (approx. 5738 KRW) per 1 million output tokens, and $0.15 (approx. 203 KRW) for cached inputs.
Meta also offers a 'Contributor' tier at lower rates: $0.10 (approx. 135 KRW) per 1 million input tokens and $0.20 (approx. 270 KRW) per 1 million output tokens, with the condition that user data may be used for Meta product improvements. Enterprises handling sensitive code or customer data should evaluate data usage terms and internal security standards alongside API costs.
Meta's release aligns with its AI open strategy, though Muse Spark 1.3 is not an open-weight model but rather available through Meta Model API and Muse Code. Open-weight models provide publicly accessible weights for users to run or adjust locally, while API-based models require interaction through company-operated interfaces.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: 100만 토큰 지원 메타 뮤즈 스파크 1.3 공개