New · October 15, 2026
Muse Spark Open
Three days after locking its frontier model behind subscriptions, Meta gave this one away: 816 billion parameters of open weights, engineered to run on a desk full of Mac Studios. Here's what you can actually do with it.
What changed
⚖️ The weights are yours
816B total parameters (42B active per token), 1M-token context, natively multimodal — download free from Hugging Face, SageMaker AI, Google Cloud Model Garden or Microsoft Foundry Models. Meta says it's the most capable US open-weight model to date.
💻 The desktop bet
Meta partnered with Apple: two M5 Ultra Mac Studios with MLX pool ~1TB of unified memory and "behave like one supercomputer." Mark Zuckerberg posted the setup from his desk and thanked Apple's CEO. Local AI that doesn't depend on the cloud — in theory.
💸 The real cost is hardware
Full-precision weights are ~1.6TB, the NVFP4 version ~530GB. The 512GB Mac Studio you need hasn't shipped yet; the 256GB one is already $10,000+. For most people, the hosted APIs (Groq, Fireworks, Together, OpenRouter…) are the sane route.
What Meta actually demoed
Meta ran Muse Code powered by Spark Open through end-to-end on-device agentic workflows: shipping real iOS changes (bug fixes, new features, full build) to Ice Cubes, an open-source Mastodon client — and driving DaVinci Resolve frame-by-frame to find the strongest moments and crop them vertically. It's a coding/agent model, not a chat toy.
How fast is it on Mac Studios?
Apple's Machine Learning Research team published benchmarks the same day the weights dropped, testing the 4-bit NVFP4 version at an 8,192-token context:
| Setup | Time to first token | Generation speed |
|---|---|---|
| 2 × M5 Ultra Mac Studios, pipeline | 8.4s | ~15.7 tokens/sec |
| 2 × M5 Ultra Mac Studios, tensor | 8.2s | ~25.0 tokens/sec |
| 4 × M5 Ultra Mac Studios, tensor | 4.8s | ~36.7 tokens/sec |
Source: Apple Machine Learning Research, Oct 15, 2026. "Pipeline" splits the model's layers across machines; "tensor" splits computation within each layer. These are decode benchmarks, not a measure of how fast an agent finishes a real project.
The hardware bar is exacting, per Meta's hardware guide and Apple's post: two Mac Studios with M5 Ultra and 512GB of unified memory each (or four with 256GB), macOS 26.2 or later, MLX 0.30.1 or later, a direct Thunderbolt 5 cable between every pair of Macs — Apple used two cables per pair to double the link speed — and RDMA switched on. In 4-bit form the weights take about 493GB, and with working memory on top Apple says a single 512GB machine is not enough. Peak draw per Mac Studio during inference is about 300W, so the rig fits in a normal office — the hard part is the price tag, not the power bill.
5 prompts that use open weights well
Written for what open weights are actually good at: private codebases, long agents, hardware you control.
Private codebase review
I can't send this code to any cloud API. Load my repo (~300 files), review it as a senior engineer, and report: the 5 riskiest functions, dead code worth deleting, and any secrets or credentials accidentally committed. Don't fix anything yet — just the report.
The Ice Cubes loop: fix a real bug
Take this open-source project's GitHub issue (#describe it), reproduce the bug locally, write the fix, add a regression test, and run the full build. If the build breaks, fix it and tell me what went wrong. Stop before pushing anything — show me the diff first.
Whole-repo archaeology
I inherited a 500k-line codebase with no docs. Map the architecture for me: what each major module does, how data flows between them, and where the most important business logic lives. Give me a one-page map I can hand to a new hire.
Frame-by-frame video assistant
I have a 20-minute talking-head recording. Watch it end to end, find the 5 strongest moments worth clipping, and tell me the exact timestamps plus a 10-word hook for each. Then suggest the crop settings for a vertical version.
Fine-tune planning
I want to adapt an open-weight model to my company's support tickets (10k examples, mixed quality). Before I burn GPU time: assess the data quality, suggest filtering rules, pick a quantization level for my 2xA100 setup, and estimate whether full fine-tuning or LoRA makes more sense. Be honest about where this will fail.
Good to know
- The Muse Model License is not MIT/Apache: free to download, but commercial terms are limited — Meta's license says companies with over $20M in yearly revenue that offer the model via API need a separate agreement with Meta. Read the license before shipping a product on it.
- Spark Open is the open cousin of Muse Wonder — Wonder stays closed, API + subscription only. Spark Open is the one you can self-host.
- Third-party quantizations go smaller than Meta's: developer reports mention ~2-bit formats running on a 256GB machine with llama.cpp — quality trade-offs apply.
- Meta says safety evaluations ran under its Superintelligence Scaling Framework, with a detailed report to follow. Treat that as pending.
- New here? Set up Muse in 15 minutes → · Get 1B free tokens →
← Muse Wonder · Muse for Windows → · Where Muse is blocked · All news
Not affiliated with Meta. Figures: Reuters, TechCrunch, MacRumors, Meta blog post, Apple ML Research benchmarks, Oct 2026.