Meta Superintelligence Labs shipped Muse Spark 1.3 (9/2), an agentic-coding iteration of the Muse Spark line: better long-horizon task sustainment, generating its own context from messy or conflicting sources, proactively correcting plan gaps, and self-tracking learnings; collaboration behaviors (clarifying questions on ambiguous prompts, asking for help when stuck, confirming before consequential actions); fewer dropped constraints on long-form instructions; better routing inside messy single-threaded contexts including interruptions; improved self-awareness of knowledge limits ('hitting hurdles rather than hallucinating outcomes'); and cleaner coding output with fewer unnecessary turns. Meta engineer comparisons vs 1.2: ~20% fewer tool calls and ~25% fewer tokens. Available in Muse Code and the Meta Model API (max-reasoning variant live on both); a Muse Spark open-weights release is teased but undated.
Meta releases Muse Spark Open for local and hosted inference
Meta released the gated Muse Spark Open weights, fulfilling the earlier open-weight preview. The 816B-parameter mixture-of-experts model activates 42B parameters per token, accepts text, images, audio and video, and has a one-million-token context window. Reference paths include Hugging Face endpoints, SGLang, vLLM, llama.cpp and oMLX. The custom license requires attribution and a separate agreement for model-as-a-service businesses above $20 million in annual gross revenue.
Availability: Muse Code + Meta Model API (max reasoning variant live); install via curl from dev.meta.ai; open weights teased, no date
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For coding-agent teams, Muse now moves from a hosted-only option to inspectable weights and multiple deployment paths. The model is still exceptionally large, gated and license-constrained, so a focused endpoint or quantized trial is more realistic than immediate self-hosting.
Evidence & related material
Meta released gated 816B-A42B weights, one-million context, a hosted endpoint and multiple local runtimes under a custom license with MaaS conditions.
Subsequent changes
Muse Spark moves from an open-weight preview to downloadable and hosted access
Technical details & scores
| Agentic | long-horizon sustainment; self-generated context from messy/conflicting sources; proactive plan-gap correction; self-tracked learnings |
|---|---|
| Collaboration | clarifying questions on ambiguous prompts; requests for help when stuck; confirms before consequential actions |
| Other | fewer dropped constraints on long-form instructions; multitasking in messy single-threaded contexts incl. interruptions; self-awareness of knowledge limits; cleaner coding, fewer turns |
| Efficiency | Meta engineer comparisons vs 1.2: ~20% fewer tool calls, ~25% fewer tokens |
| Availability | Muse Code + Meta Model API (max reasoning variant live); install via curl from dev.meta.ai; open weights teased, no date |
| Openness | API/tool access only at release; no weights yet |
| Open Release | gated Muse-Spark-Open-816B-A42B weights |
| Architecture | 816B total / 42B active MoE; 1M context |
| Modalities | text, image, audio and video input; text output |
| Deployment | Hugging Face endpoints, SGLang, vLLM, llama.cpp and oMLX; BF16 reference needs 4x8 H200, NVFP4 one B200 node |
| License | custom Muse Model License; attribution; separate agreement for MaaS businesses above $20M annual gross revenue |