Meta's Muse Spark Open: 816B Parameters You Can Download Today (But Probably Can't Run)

Meta launched Muse Spark Open on October 15, 2026 — an open-weight model it calls the "most capable U.S. open-weight model to date." The weights are free to download right now. The catch: running it locally needs about a terabyte of RAM.

Here's what developers need to know.

What is Muse Spark Open?

Muse Spark Open is the open-weight sibling of Muse Spark, the model behind Meta AI. It launched three days after Meta's closed frontier model, Muse Wonder.

The headline specs:

• 816 billion total parameters (mixture-of-experts, only 42B active per token)
• 1 million token context window (1,048,576 tokens)
• Natively multimodal: reads text, images, audio, and video; writes text
• Built for agentic coding: long multi-step tasks, tool use, real bug fixing

Two checkpoint formats ship from Meta: full-precision BF16 (~1.6TB) and a compressed NVFP4 version (~530GB). Community quants (AMD, Intel, Unsloth) have produced smaller formats down to roughly 2-bit that reportedly run on a 256GB machine with llama.cpp.

How do you actually use it?

Three routes, depending on your hardware budget:

1. Download the weights (free)
Available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Free to download with a license allowing limited commercial use.

2. Run it locally (expensive)
Meta worked with Apple to optimize the model for the MLX framework. It runs on two 512GB M5 Ultra Mac Studios linked together — Apple's unified memory architecture pools them into what Meta calls "one supercomputer with a terabyte of unified memory." The 512GB Mac Studio isn't even shipping yet (expected later this month), and the 256GB version already costs over $10,000.

3. Use a hosted API (practical)
Day-zero launch partners include Together AI, Baseten, Fireworks AI, Groq, Modal, Nebius, and OpenRouter. Together AI offers both serverless and dedicated infrastructure.

What is it actually good at?

Per Meta's own release and early-access partner reports (vendor claims, not independently verified):

• Long-horizon agentic coding: keeping entire codebases in context, planning, acting, observing, and self-correcting over extended tasks
• Tool calling: Ollama's early testing reported strong performance on multi-step and parallel tool calls with clean structured outputs
• Multimodal understanding: documents, images, audio, and video in a single prompt — Meta demoed it reviewing and cropping footage in DaVinci Resolve
• Commentary mode: the model states what it is doing and why before acting — useful for agent observability

The catch

This is the section that matters.

The hardware bar is absurd. One terabyte of unified memory puts local inference firmly in "company workstation" territory, not "developer laptop." The 2-bit community quants on a 256GB machine are interesting but unverified at scale.

The license has teeth. The Muse Model License requires prominent "Built with Muse Spark Open" attribution. Companies over $20 million in revenue must notify Meta and sign a commercial agreement before offering it as a service. This is not MIT-licensed freedom.

Benchmarks are mixed. Meta's benchmarks show strong agentic and tool-use performance — but weaker coding scores versus some leading Chinese open models. "Most capable U.S. open-weight model" is a carefully worded claim: it sidesteps the global competition.

No independent benchmarks yet. Every performance number in this article comes from Meta or its launch partners. Treat accordingly.

Who should care, and who shouldn't

Worth your time if you:

• Build agentic coding workflows and want an open-weight model with a 1M context window
• Need multimodal input (screenshots, voice notes, screen recordings) in one model
• Already run Together AI, Fireworks, or Groq infrastructure

Skip it if you:

• Wanted a model you can run on your laptop (you can't, unless you have $20k+ in Mac Studios)
• Need a permissive commercial license without attribution strings attached
• Already have Qwen or Kimi-class open models working for your coding tasks

TL;DR

• Meta released 816B-parameter open weights on October 15, 2026 — free download
• 42B active params, 1M context, multimodal input, built for agentic coding
• Local run needs ~1TB RAM (two M5 Ultra Mac Studios); API access via 7 launch partners
• License requires attribution; $20M+ revenue companies need a commercial agreement
• Strong agentic benchmarks per Meta, weaker coding scores vs Chinese open models — all vendor claims so far

FAQ

Is Muse Spark Open really free?
The weights are free to download and the license allows limited commercial use. But attribution is mandatory, and larger companies need Meta's sign-off.

Can I run it with Ollama?
Ollama was a day-zero evaluation partner and reported good tool-calling results. Check their model library for available quants — full weights need far more RAM than a typical machine has.

How does it compare to Muse Wonder?
Wonder is Meta's closed frontier model (launched October 12). Spark Open is the open-weight release of the Spark family — related lineage, different access model.

What's the cheapest way to try it?
A hosted API endpoint (Together AI, Fireworks AI, Groq, or OpenRouter) — no hardware investment, pay per token.

---

Sources: Meta's October 15 announcement (via TechCrunch, MacRumors); Together AI launch statement; community analysis of checkpoint formats. All capability and benchmark claims are vendor-reported unless noted.