Meta spent four days doing two opposite things. On October 12 it released Muse Wonder and Muse Wonder Full, closed models that put it near the top of the independent leaderboards. On October 15 it published the weights of Muse Spark Open, an 816-billion-parameter mixture-of-experts model, roughly twice the 405 billion parameters of Llama 3.1, Meta's 2024 open flagship. Meta now runs a closed frontier business and an open-weights one side by side, and the most revealing document of the week is not a benchmark table but a license.
The closed half sits near the top
On the Artificial Analysis Intelligence Index, Muse Wonder Full scores 56, tied for second overall, two points behind Claude Opus 5.5 and level with Claude Sonnet 5.5. Muse Wonder scores 52, seventh overall, up from 48 for its predecessor, Muse Spark 1.3.
The price moved with the score: $4 per million input tokens and $20 per million output, against $1.25 and $4.25 for Spark 1.3. Per index task, Wonder costs $4.52, against $1.60 for Spark 1.3 and $5.98 for Claude Opus 5.5. Wonder Full has no API price at all. It lives inside Muse Code, Meta's coding agent, on a subscription The Batch lists at $50 a month, with a $200 tier reported by NewsBytes. The best model is a reason to subscribe, not a line item.
The weak spot is recall. On AA-Omniscience, Artificial Analysis's knowledge test, Wonder Full scores 30 against 46 for Claude Opus 5.5, though its hallucination rate fell from 33% to 23%. It knows less and admits it more.
The open half: 816 billion parameters, 42 billion at a time
The model card lists 816 billion total parameters, 42 billion active per token, and 256 experts with eight routed per token, in BF16 and NVFP4 checkpoints. Artificial Analysis scores it 39: first among US open-weights models, seventh among open-weights models overall, behind MiMo-V2.6-Pro and GLM-5.3. Meta's own tables locate the gap in coding: on Terminal-Bench 2.1 it scores 76.4, against 85.0 for Kimi K3 and 83.9 for GLM 5.3. Weights and hosted APIs went live on launch day across Hugging Face, the three big clouds and seven API providers.
Note what Meta opened. Spark Open sits nine points below the closed Spark 1.3 and 17 below Wonder Full. Meta gave away the model it no longer needs to sell.
The hardware bill, in bytes
At two bytes per parameter the checkpoint comes to about 1.63 TB, and the model card asks for at least 2 TB of GPU memory for BF16 and at least 700 GB for NVFP4. Hugging Face's guide spreads the BF16 weights across four nodes of eight H200s and fits the NVFP4 weights on a single B200 node.
Why 700 GB, when 816 billion parameters at four bits come to about 408 GB (our arithmetic)? Context. Hugging Face puts the KV cache at roughly 250 KB per token, about 266 GB for a full million tokens. The weights are the down payment; the context window is the mortgage. The Batch's estimate of about 1 TB at 4-bit runs high by that math, and lands instead on the combined memory of Meta's showcase rig: two 512 GB Mac Studios with M5 Ultra chips, linked with MLX. Apple does not yet sell that configuration to consumers, The New Stack notes, putting the idea at roughly the "price of a compact car."
There is a cheaper floor. Hugging Face ran a 2-bit quant on one 512 GB M3 Ultra at about 18 tokens per second in roughly 304 GB, and Unsloth lists 512 GB for its 4-bit build and 870 to 950 GB for 8-bit. Speed follows the active slice: at four bits each token reads about 21 GB of weights (our arithmetic), so memory bandwidth, not capacity, sets the pace.
Open, by whose definition
The road here is short. Muse Spark arrived on April 8 as the first model from Meta Superintelligence Labs and Meta's first without open weights, a break Bloomberg also noted. The $299 Meta Glasses that run it followed on June 23, an API with Spark 1.1 in July, and in September Zuckerberg promised open weights "soon". Soon was October 15.
The Muse Model License is generous where most users live. Internal use, research and building the model into your own product are free, and Meta claims no ownership of outputs. Everyone must prominently display "Built with Muse Spark Open" and follow a use policy that bars military uses, weapons and the operation of critical infrastructure. And a company selling the model as a service must notify Meta and sign a separate agreement once its group revenue passes $20 million in any 12-month period. The Batch read the threshold as a general commercial cap; the license text confines it to model-as-a-service businesses.
Hold that against the Open Source Initiative's definition, which since October 2024 requires that an AI system can be used for any purpose without asking permission. A permission gate for one class of business and a list of forbidden fields fail that test, on the grounds OSI used when it called Llama's license "very plainly not an 'Open Source' license", a verdict it restated in 2025. That is our reading of OSI's criteria, not an OSI ruling. Meta, to its credit, calls the model open-weight.
Against Llama's history, the strings moved rather than loosened. Llama 3.3's license required a separate license above 700 million monthly active users and a "Built with Llama" label. The attribution habit survived; the ceiling fell from a number that touched a handful of giants to a revenue line that catches any mid-sized inference host. The New Stack found the terms close to Moonshot's license for Kimi K3, while Reflection has promised plain Apache 2.0 for its 501B Beam.
The 30,000-foot read
Meta has stopped choosing between open and closed and started pricing the difference. The frontier stays behind an API and a subscription; the runner-up goes out as weights, where it earns developer loyalty, cloud distribution and a credible American answer to Qwen, GLM and Kimi. The license is the hinge: free for the many who would never have paid, gated for the few who could undercut the API at scale. It is a coherent strategy, and it is not open source.
What you control is narrower: whether 700 GB of accelerator memory buys you more than an API key, and whether you can live with a label on your product and a call to Menlo Park past $20 million. If Spark Open is the pattern, Meta's open models will trail its closed ones by a generation and a license. If Meta ever opens a Wonder, read the license before the benchmarks.
Sources: Muse Spark Open model card, Hugging Face (primary: parameters, experts, October 15 release, checkpoints, VRAM guidance, license and use-policy terms, benchmark table); Hugging Face blog, "Welcome Muse Spark Open" (H200 and B200 configurations, KV-cache size, M3 Ultra test); The Decoder on Muse Wonder (October 12 launch, pricing, index scores, per-task costs, knowledge and hallucination results); The Decoder on Muse Spark Open (positioning against Qwen, GLM and Kimi); The Batch, issue 376 (index rankings, $50 Muse Code plan, 1 TB estimate, broader license summary, Llama 3.1 size); The New Stack (launch-day availability, Meta's Terminal-Bench 2.1 table, Mac Studio setup and consumer availability, open-weight wording, Kimi K3 license comparison); Unsloth documentation (checkpoint size, memory by quantization); NewsBytes ($200 Muse Code tier); Simon Willison on Muse Code and Muse Spark 1.2 and on Muse Spark 1.1 (Muse Code as Meta's coding agent, July API); The Decoder on Muse Spark and Bloomberg (Muse Spark's April 8 debut as a closed model from Meta Superintelligence Labs); The Next Web ($299 Meta Glasses running Muse Spark, June 23); The Register (Zuckerberg's September open-weights promise); Gizmodo (OSI's Open Source AI Definition 1.0); OSI, 2023 and OSI, 2025 (Llama license verdicts); Llama 3.3 Community License, Wind River copy (700 million MAU clause, attribution); Reflection (Apache 2.0 plan for Beam). CAVEATS and corrections: Meta's research blog post refused automated retrieval, so Muse Wonder figures come from The Decoder and The Batch rather than from Meta directly; Spark Open's specs and license terms were checked on the Hugging Face model card. Correction: early notes placed Muse Spark's debut in June 2026; it launched April 8, 2026, and the $299 glasses that run it followed on June 23. Correction: the $20 million threshold applies only to model-as-a-service businesses per the license text, not to all commercial users as The Batch's summary suggests. The "about 1 TB at 4-bit" figure is The Batch's; the model card asks for at least 700 GB for NVFP4, and 408 GB of 4-bit weights plus a full-context KV cache is our arithmetic. The $200 Muse Code tier is single-sourced to NewsBytes; The Batch lists only the $50 plan. Terminal-Bench 2.1 results are Meta-reported; index scores are Artificial Analysis's, via The Batch and The Decoder. The model card's metadata lists 822B parameters against 816B in its text. The OSI comparison applies OSI's published criteria; we found no OSI ruling on the Muse Model License. The Llama 3.3 license was read through Wind River's copy because Meta's license page refused retrieval.