Meta Muse Spark Open: 816B open weights on two Mac Studios
Meta Muse Spark Open is an 816B open-weight coding model that runs on two M5 Ultra Mac Studios. Read the Muse Model License before you self-host.
Meta shipped Muse Spark Open on October 15. It is an 816-billion-parameter open-weight model built for coding and agent work, and Meta says it runs on two Mac Studios on a desk. That is a real shift for local AI. It is also not "free Llama" again. The weights come with a license, and the license is where a business should start.
What actually happened
Per Meta's announcement and TechCrunch:
- Size: 816B total parameters, 42B active per token (mixture-of-experts)
- Context: 1 million tokens, natively multimodal
- Focus: long-horizon coding, software engineering and multi-step agent tasks; Meta paired it with its Muse Code terminal
- Hardware: two Apple M5 Ultra Mac Studios with MLX. Each has up to 512GB of unified memory, so about 1TB total
- Weights: Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry
- Hosted APIs: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI
- Engines: vLLM, SGLang, TensorRT-LLM, llama.cpp, MLX and others
- License: the Muse Model License, which Meta describes as having "limited commercial terms"
Meta calls it the most capable US open-weight model to date. Its announcement did not give benchmark numbers in the text we could read, and the full evaluation report is "available soon." Treat the ranking as a vendor claim until third-party tests land.
Why Muse Spark Open matters for your business
Local frontier-class coding is now a hardware purchase. Two Mac Studios is a capital expense, not a data center. For a shop with sensitive client code, that changes the math against a per-token API bill. Run the numbers on your own token volume before you buy.
"Open weights" is not "open license." Meta says the terms let "many developers" use the model for free. Many is not all. The thresholds live in the model card. Read them before you build a product on it, and save a copy of the version you agreed to.
42B active is the number to watch. Speed depends on the active parameters, not the 816B headline. That is why it fits on desktop silicon at all. Test it on your real repo, not a demo.
Start hosted, then decide. Seven API providers carry it today. Run your eval set through OpenRouter or Together first. Buy hardware only if the results and your volume both justify it.
Key takeaways
- Muse Spark Open: 816B total, 42B active, 1M-token context, coding and agent focus
- Meta says it runs on two M5 Ultra Mac Studios with about 1TB of unified memory
- Weights ship under the Muse Model License with limited commercial terms
- Benchmark claims are Meta's own until the full eval report and outside tests land
- Test it through a hosted API on your own tasks before you buy hardware
A model you can host is only portable if your stack can swap it in. We build vendor-agnostic AI systems that run on hosted APIs or your own hardware behind one interface, with an eval set to prove the swap. See how we build it, or check whether local hardware beats your API bill.
Sources: Meta, TechCrunch.
- #meta
- #open-weights
- #local-ai
- #mac-studio
- #self-hosting
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Floqer AI agents keep sales CRM data fresh, open source in 2027
Floqer's AI agents research sales prospects across 100+ data sources and plan to open-source the customer database. Fix your CRM data before you buy prospecting AI.
Read itX Subscription bundles Grok, Cursor and X Premium: $8 to $200
SpaceXSI's X Subscription bundles Grok, Cursor and X Premium into one plan from $8 to $200 a month. Bundles are cheap until you need to leave one piece.
Read it