Muse Spark 1.3 (Meta)

This TrueFoundry article confirms that Meta Superintelligence Labs released Muse Spark 1.3 on September 2, 2026, roughly four weeks after version 1.2, as a natively multimodal model targeted at agentic and coding work, available through Muse Code and the Meta Model API. It states Muse Spark is the first model family fr The article reports Meta's internal claim that Muse Spark 1.3 uses roughly 20% fewer tool calls and 25% fewer tokens than 1.2 on the same tasks, and notes additional changes including a max reasoning level, better multitasking inside one long thread, and stronger resistance to prompt injection. It also cites Meta Model

This DataCamp tutorial covers Meta's Muse Spark 1.3 release and reports that the model shipped on September 2, 2026 from Meta Superintelligence Labs as a flagship multimodal reasoning model with a 1,048,576-token context window that accepts text, images, video, and files. Meta's own claim, cited by the tutorial, is tha The tutorial also walks through hands-on testing of Muse Code, finding across six runs on three coding tasks that Muse Spark 1.3 was cheaper on two tasks and 38% more expensive on the third, netting a 12% cost increase over the set, while model completions fell 23% and 32% on the two tasks where 1.3 won and uncached in

This MarkTechPost article reports that Meta Superintelligence Labs released Muse Spark 1.3, described as the fourth Muse Spark release in five months and aimed at long-horizon agentic and coding work rather than single-turn generation, framing the update around usability, sustained threads, user collaboration, and reco It also reports Meta's claimed efficiency gains of roughly 20% fewer tool calls and 25% fewer tokens compared with Muse Spark 1.2, and explains that Meta trained the model across multiple agent harnesses so behavior generalizes past one environment. The model is built to hold several workflows inside a single long thre

Meta released Muse Spark 1.3 on September 2, 2026 and began rolling it out the same day through Muse Code and the Meta Model API. The checkpoint targets longer agentic and coding tasks, with Meta reporting fewer tool calls and lower token use than Muse Spark 1.2. Muse Spark 1.3 is a closed-weights reasoning model that The article reproduces Meta's benchmark scorecard comparing Muse Spark 1.3 against Claude Opus 5: DeepSWE v1.1 75.4 vs 74.0; SWEAtlas CodeBase QnA 59.4 vs 52.7; MRCR 512K–1M recall 98.1 (Opus 5 has no published figure); GDPval-AA v2 1754 vs 1824; OSWorld 2.0 66.9 vs 68.3; and AutomationBench 49.4 vs 50.3. An important

Meta launched Muse Spark 1.3 on September 3, 2026 with a focus on coding, agentic workflows, and complex multi-step tasks. According to Meta AI Research, the model achieves roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 during coding tasks. Meta says the model is designed to sustain longer-runnin The release also emphasizes more collaborative behavior: Muse Spark 1.3 can ask clarifying questions when instructions are ambiguous, request assistance when stuck, and seek confirmation before consequential actions. It is also tuned to follow lengthy and complex instructions without losing individual requirements and

Meta released Muse Spark 1.3 on September 3, 2026, replacing Muse Spark 1.2 as the model behind Muse Code. According to Meta's own benchmark table, the model leads or ties Claude Opus 5 and GPT-5.6 Sol on two coding benchmarks: DeepSWE v1.1 (75.4 vs Opus 5's 74.0 and GPT-5.6 Sol's 73.0) and Terminal-Bench 2.1 (88.8, ty Two pricing tiers distinguish the rollout: the standard muse-spark-1.3 at $1.25/M input and $4.25/M output (data not used for training) and a contributor tier at $0.10/M input and $0.20/M output in exchange for Meta training on usage data, a roughly 12.5x input and 21x output discount. Mark Zuckerberg announced the lau