Muse Spark Open
Released on October 15, 2026
Changes from previous version
Publishes downloadable weights after hosted Muse Spark 1.3. Context stays at 1,048,576 tokens, and the hosted token rate matches 1.3 at $1.25/$4.25 with cache reads at $0.15. Meta's xhigh scorecard reports Terminal-Bench 2.1 at 76.4% with Terminus 2 and DeepSWE v1.1 at 60% on the best available harness. Measured price per task is $0.84 at xhigh.
Release Summary
Meta's open-weight Muse Spark model for coding and long-horizon agents. 816B total parameters, 42B active, a 1,048,576-token context, and text, image, video, and audio in. models.dev lists an output cap of 131,072 tokens. BF16 and NVFP4 weights ship under the Muse Model License, and the model runs on two Apple M5 Ultra Mac Studios through MLX. Hosted APIs at Together AI, Fireworks, Baseten, Modal, Groq, Nebius, and OpenRouter list $1.25/$4.25 per million tokens, with cache reads at $0.15. Meta does not publish a first-party API price. On Meta's xhigh scorecard, JobBench is 61.8, MCP-Atlas is 89.2, Terminal-Bench 2.1 is 76.4% with Terminus 2, and DeepSWE v1.1 is 60% on the best available harness. Measured price per task at xhigh is $0.84. Hosted predecessor: Muse Spark 1.3.
Timeline
Muse Spark Open released
Meta publishes Muse Spark Open weights and hosted inference at $1.25 per million input tokens and $4.25 per million output tokens.