Meta released the closed frontier model Muse Wonder on October 12, 2026, followed three days later by the open-weight release of Muse Spark Open on Hugging Face, according to OrcaRouter. Rather than serving as direct versions of the same software, Muse Wonder operates as a proprietary subscription and API service, while Muse Spark Open delivers 816 billion mixture-of-experts parameters via an accessible model download and license.
What separates Muse Spark Open from Muse Wonder
The release schedule set the two systems apart from their debut. Meta issued Muse Wonder first in two checkpoints: “muse-wonder-1.0” through the Model API and inside Muse Code, alongside “muse-wonder-full-1.0,” which is sold strictly via a Muse Code subscription, OrcaRouter reported.
Muse Spark Open arrived on October 15, 2026, packaged as a single artifact containing a BF16 checkpoint, an NVFP4 build behind an access-controlled repository, a technical report, and the Muse Model License. One functions as a managed service, while the other functions as local files requiring dedicated hardware and serving infrastructure.
Both models process text, images, voice, and video inputs to generate text outputs, and both share a declared context window of 1,048,576 tokens. Beyond input modalities and context limits, their economic frameworks and technical architectures diverge completely.
How Artificial Analysis scored their intelligence
Artificial Analysis evaluated both models under its updated Intelligence Index v5.0.1, incorporating nine distinct benchmarks including AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2, and Terminal-Bench-Science 0.1, OrcaRouter reported.
On that evaluation suite, Muse Spark Open running at Xhigh effort scored 25 against a class median of 23 for similarly sized open-weight models. Muse Wonder at Max effort achieved a score of 44 compared to a median of 40 for models in its respective price tier.
Individual benchmark metrics highlight distinct performance gaps across agentic and technical workloads. On Terminal-Bench 4.0, which tests command-line execution agents, Muse Spark Open scored 0.081 while Muse Wonder reached 0.520. For Terminal-Bench-Science 0.1, Muse Spark Open recorded 0.019 against 0.257 for Muse Wonder.
The narrowest margin appeared in knowledge accuracy on AA-Omniscience v1.1, where Muse Spark Open scored 25.8 and Muse Wonder scored 25.1. However, hallucination rates diverged in that pairing, with Muse Spark Open registering 28.0% compared to 20.9% for Muse Wonder, OrcaRouter reported.
Why parameter counts dictate hardware requirements
Muse Spark Open publishes its structural dimensions openly, totaling 816 billion parameters with approximately 42 billion active parameters per token. The architecture features 62 decoder layers and 256 mixture-of-experts components per layer, selecting eight experts per token.
The distributed quantized build keeps experts in 4-bit NVFP4 alongside select layers retained in BF16. The 42 billion active parameter count dictates serving constraints, allowing the model to run on eight GPUs.
By contrast, Muse Wonder does not disclose parameter counts in verified technical documentation, with Artificial Analysis designating the architecture as proprietary. While specific parameter totals remain undisclosed, its pricing tier and Index score indicate a substantially larger underlying scale managed entirely via external infrastructure.
What speed and output volume reveal about reasoning
Execution speeds and token generation rates differ across both hosted endpoints. Muse Spark Open yielded an output speed of 169 tokens per second with a time-to-first-token of 13.81 seconds, while Muse Wonder delivered 118 tokens per second with a time-to-first-token of 25.19 seconds, OrcaRouter reported.
Task completion timing followed a similar divergence. An Index evaluation task required an average of 6.6 minutes for Muse Spark Open and 13.6 minutes for Muse Wonder. Muse Wonder consumed 113 million output tokens during the evaluation run, whereas Muse Spark Open utilized 90.6 million output tokens.
How real-world operating costs shift the calculus
Calculated cost per Index task accounts for total evaluation suite expenses divided by task volume. Muse Spark Open incurred an average cost of $2.02 per task, while Muse Wonder averaged $6.95 per task, according to OrcaRouter data.
List pricing structures show blended rates of $2.00 per million tokens for Muse Spark Open under a 3:1 input-to-output ratio, compared to $8.00 per million tokens for Muse Wonder. Caching discounts alter these figures further, with Muse Spark Open offering an 88% discount at $0.15 per million tokens and Muse Wonder offering a 95% discount at $0.20 per million tokens.
Licensing terms introduce additional compliance considerations for commercial deployments. The Muse Model License for Muse Spark Open stipulates that if third-party access to the model via an API generates a gross annual revenue exceeding $20 million over any consecutive 12-month period, developers must notify Meta and secure a separate written agreement.
Keep reading
