Meta had a strange October. On October 12, it launched Muse Wonder — its first frontier-class AI model — and locked the weights away completely. Three days later, it released Muse Spark Open, an 816-billion-parameter monster, and gave the weights away for free. Same company, same month, opposite doors. So which one is the real Meta?
I've read Meta's announcement, the model card, and the press coverage (TechCrunch's launch piece is the cleanest summary). My honest read: Spark Open is genuinely interesting, the "free" part comes with fine print, and the locked-door Wonder launch says more about where Meta is heading than the open release does.
What Muse Spark Open actually is
Strip the marketing: Muse Spark Open is a natively multimodal mixture-of-experts model — 816 billion total parameters, 42 billion active per token, and a 1-million-token context window. It reads text, images, video, and audio; it writes text. The weights are free to download from Hugging Face, Amazon SageMaker AI, Google Cloud, and Microsoft Foundry.
Meta built it for coding and agentic work: long multi-step tasks, tool use, messy real-world codebases. The launch demos were uncharacteristically honest — fixing real bugs in the open-source Ice Cubes Mastodon client, and driving DaVinci Resolve to review footage and cut a vertical clip. I respect that.
The benchmark story needs the usual discount. Meta calls it the "most capable U.S. open-weight model to date" — note the "U.S." doing heavy lifting there, because the strongest open-weight competition right now comes from Chinese labs (DeepSeek, Qwen, Kimi, GLM). Meta's scorecard compares Spark Open against models like Mistral 4 Large, Qwen3.8-2.4T-A95B, Kimi K3, and GLM 5.3. And buried in the fine print, VentureBeat spotted that the model card cites an Artificial Analysis Intelligence Index score of 39 — below MiMo-V2.6-Pro's independently published 46. Strong, not dominant. Vendor charts are a ceiling, not a floor.
The Mac angle is real — and mostly theater
Meta worked with Apple to get Spark Open running on Mac Studio hardware via the MLX framework. The catch: it takes two 512GB M5 Ultra Mac Studios (or four 256GB ones), wired together, to pool roughly a terabyte of unified memory. Meta reports about 25 tokens per second on two machines and about 37 on four, in 4-bit precision.
Here's the funny part: Apple hadn't even shipped the 512GB Mac Studio at launch, and the 256GB version already costs over $10,000. Real engineering, absurd practicality. For nearly everyone, the actual routes are the hosted APIs — Together AI, Fireworks AI, Groq, Baseten, Modal, Nebius, and OpenRouter all launched with it — or downloading the weights yourself. The full-precision checkpoint is about 1.65TB; the compressed NVFP4 version is around 530GB. "Free to download" is accurate. "Free to run" very much is not.
The license is free — with strings attached
This matters more than the benchmarks. Spark Open ships under the Muse Model License, a custom Meta license — not an open-source license in the Apache/MIT sense. Most developers can use it for free, including commercially. But three terms matter:
- Attribution. Anything you build on it has to display "Built with Muse Spark Open."
- The $20 million trigger. If your company makes more than $20 million a year and wants to sell access to the model as an API, you have to notify Meta and sign a separate deal first. Using it inside your own product, for internal work, or for research doesn't trigger this.
- No training code, no dataset. Weights only. This is "open" the way a recipe book without the ingredients list is open.
My read: Meta wants the ecosystem effects of open models — developers building, researchers poking, everyone talking — without funding a competitor's API business. That's not evil. It's just not the Llama-era generosity people remember. It's a business decision wearing an open-source costume. Compare it with my DeepSeek V4.1 Flash review, where the free model genuinely costs almost nothing to actually use.
Three days earlier: Muse Wonder stayed locked
Now the other door. On October 12, Meta Superintelligence Labs — the org led by former Scale AI CEO Alexandr Wang — launched Muse Wonder and Muse Wonder Full, Meta's first frontier models. Neither ships with downloadable weights. VentureBeat put it bluntly: only the weaker one is available via API, and even that one isn't downloadable.
The access rules are a maze. Standard Wonder is on the Meta Model API at $4 per million input tokens and $20 per million output (cached input is $0.20) — priced to the cent exactly like Anthropic's Claude Opus 5.5. Wonder Full, the bigger model that wins Meta's best benchmark rows, isn't on the API at all. It lives only inside Meta's Muse Code terminal agent, on $50 or $200-a-month plans. (How the same market gets priced from the consumer side: my ChatGPT Plus breakdown.)
On independent numbers, Artificial Analysis scores Wonder at 52 and Wonder Full at 56 on its Intelligence Index — behind Opus 5.5's 58. So Meta's most powerful model ever is a tier below the best, priced the same as the best, and you can't download it — or even call the best version through an API. Meta's stock still jumped about 3.4% that day.
So what is Meta actually doing?
Here's my take, and I think it's the part most coverage misses: this is a hedge, not a contradiction.
Meta's Llama models taught the whole industry that open weights build ecosystems. But ecosystems don't pay for $89 billion of capital spending — Meta's trailing four-quarter capex, per TIKR. Frontier models are now too expensive to give away. So Meta split the difference: give away the workhorse, rent the crown jewel. Spark Open keeps developers, researchers, and the open-source conversation inside Meta's orbit. Wonder — priced to match Opus 5.5 — is how Meta starts getting a return on the data-center bill.
The thing I'd actually watch: whether Meta opens Wonder's weights a year from now, once the next frontier model arrives. That's been the industry's pattern. Zuckerberg promised Spark's open weights about six weeks before shipping them. Nobody has promised anything for Wonder.
Which door should you walk through?
Practical advice, no fluff:
- If you're a developer or small team building agentic coding tools: try Spark Open through a hosted API (OpenRouter or Together AI are the easy doors). Near-frontier coding without the per-token bill, self-hostable later. The attribution line is a small price.
- If you're a business choosing a foundation model: the real question isn't open vs. closed — it's who pays the compute bill and who controls your roadmap. Pick your landlord consciously.
- If you're just curious: wait. The 512GB Mac Studios weren't shipping at launch, the 1.65TB download wants a data center, and API prices will move. The models aren't going anywhere.
For how agent-ready models plug into your tools, my MCP explainer is the missing piece — and the GPT-6 Astra breakdown shows OpenAI's answer at a similar tier.
FAQ
Is Muse Spark Open really free?
The weights are free to download and use, including for limited commercial use, under the Muse Model License. Running it is the expensive part: about a terabyte of RAM locally, or a hosted API bill. Companies making over $20 million a year that want to resell it as an API need a separate agreement with Meta first.
Can I run Muse Spark Open on my laptop?
No. Meta's official setup needs two 512GB M5 Ultra Mac Studios (or four 256GB ones) linked together, and the compressed checkpoint alone is around 530GB. Community builders have reportedly squeezed it onto 256GB machines with heavy quantization, but that's unofficial territory.
Is Muse Wonder open source?
No. Neither Muse Wonder nor Muse Wonder Full ships with downloadable weights. Standard Wonder is available through Meta's Model API at $4/$20 per million input/output tokens; Wonder Full is only available inside Meta's Muse Code agent on $50 or $200 monthly plans.
Which is better — Muse Spark Open or Muse Wonder?
Different tools for different jobs. Wonder (especially Wonder Full) scores higher on independent benchmarks, but you rent it per token and can't inspect or host it. Spark Open scores lower but is free, downloadable, and yours to run. Capability vs. control — decide which your project actually needs.
What does the Muse Model License require?
Attribution ("Built with Muse Spark Open"), plus a separate agreement if your company earns over $20 million annually and sells API access to the model. Internal use, research, and building it into your own product don't need that extra step.
Meta opened one door and locked another, in the same week, on purpose. Watch which one makes money — that'll tell you where the next model lands.