Meta launches muse spark 1.3: a more efficient agentic coding model

Meta Superintelligence Labs has released Muse Spark 1.3, an advanced agentic coding model designed for long-horizon tasks. The update, which ships today, is part of a series of four releases in five months. Muse Spark 1.3 is optimized for usability, enabling sustained long threads, user collaboration, and awareness of its own limitations. It is available via the Muse Code and Meta Model API, though self-hosting is not possible due to closed weights and safety testing restrictions. The model demonstrates improved efficiency, using approximately 20% fewer tool calls and 25% fewer tokens compared to Muse Spark 1.2. These reductions translate to lower costs for agentic workloads. Meta reports enhanced multitasking capabilities, with the model adept at mapping incoming prompts to the correct task within a single thread. It also shows better calibration on its own limits, flagging hurdles instead of hallucinating outcomes. Performance metrics highlight Muse Spark 1.3’s strengths, with scores of 75.4 on DeepSWE v1.1, surpassing competitors like Claude Opus 5 and GPT-5.6 Sol. Long-context retrieval shows the widest gap, with MRCR v2 scores of 98.5 and 98.1 against GPT-5.6 Sol’s 91.5 and 73.8. Artificial Analysis places the xhigh variant at 61 on its Intelligence Index, trailing Claude Opus 5’s max variant at 63.