Meta Muse Spark Open Explained: 816B Free Model (2026)

On October 15, 2026, Meta released Muse Spark Open — the weights of an 816-billion-parameter AI model that anyone can download for free and build on. (Actually running it is another matter — more on the real costs below.) Meta calls it the most capable open-weight model yet from a US lab. It lands just three days after Meta unveiled its closed flagship, Muse Wonder — and it's the one Meta chose to give away.

So what does an 816-billion-parameter free model actually mean for you? The short version: this is Meta's answer to the Chinese open-weight models (Qwen, Kimi, GLM) that have dominated downloadable AI through most of 2026 — and it's aimed squarely at developers, coding assistants and AI agents. Here's the full picture.

What Meta actually released

Muse Spark Open is a mixture-of-experts (MoE) model: it holds 816 billion parameters in total, but only about 42 billion are "active" for any single token, which keeps it efficient. The headline specs, from Meta's announcement:

• 816B total parameters, ~42B active per token
• 1-million-token context window — enough for an entire codebase or very long documents
• Native multimodal input: text, images, audio and video in; text out
• Built for coding and agentic work: long-horizon multi-step tasks, tool use and real software debugging
• Available on Hugging Face (BF16 and NVFP4 checkpoints), Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models
• Hosted APIs from seven launch partners: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI

In Meta's self-reported demos, a coding tool powered by Spark Open shipped fixes to an open-source iOS app and drove the DaVinci Resolve video editor to review footage frame by frame and crop clips to vertical format — all running locally, without cloud GPUs (Meta's claim; not independently verified).

The verification caveat: every performance number comes from Meta or its launch partners. The Artificial Analysis Intelligence Index figures cited in launch coverage (GLM 5.3 at 45 vs Spark Open's 39) come from Meta's own comparison table in its model card — Meta-selected figures, not an independent verdict. Meta labels the rival only "GLM 5.3" (some discussion refers to "GLM 5.3 Flash"), so the exact variant is ambiguous; other rows of the same table favor Spark Open (MCP-Atlas 89.2 vs 84.2 for GLM 5.3 and 82.3 for Kimi K3). No independent benchmarks exist yet — so treat "most capable US open-weight model" as Meta's own claim, not a settled fact.

The Apple twist: runs on two Mac Studios

The genuinely surprising part isn't the model — it's the partnership. Apple and Meta spent years sniping at each other after Apple's privacy changes hit Meta's ad business. Yet Meta worked directly with Apple to optimize Spark Open for Apple's MLX framework, so the weights can be split across two linked Mac Studios that behave as one machine.

The reference setup: two M5 Ultra Mac Studios with 512GB of unified memory each — roughly a terabyte of total memory, with the chip's high memory bandwidth reading the small slice of the model each token needs. Meta's Mark Zuckerberg even posted about it publicly, crediting Apple's team.

Why it matters: every demo proving serious AI work can run on Mac hardware instead of a data center full of NVIDIA GPUs is a win for Apple's silicon story. For everyone else, it's proof that desktop-class AI is moving upmarket fast — from 8GB laptops running small models to paired workstations running frontier-scale ones.

"Open-weight" is not "open source" — read the license

This matters more than most coverage admits. Muse Spark Open ships under Meta's custom Muse Model License, not a permissive license like Apache 2.0 or MIT. Three terms to know:

1. Attribution is mandatory. Any product built on the model — or on anything derived from it — must prominently display "Built with Muse Spark Open" in its user interface.
2. Model-as-a-Service trigger. If the licensee and its affiliates together exceed $20 million in aggregate gross revenue during any consecutive 12-month period, they must notify Meta and sign a separate written agreement before offering the model as a paid API service. Using it inside your own product, for internal work, or for research is not covered by that clause. Meta also claims no ownership of your outputs.
3. Training data and code are not published. You get the weights, not how they were made.
4. Usage policy restrictions. The accompanying usage policy bars military and critical-infrastructure uses and says the model is not intended for under-18s.

For freelancers, startups and researchers, this is effectively free. For larger companies reselling model access, it's a gated license. Calling it "open source" without qualification is wrong.

How to actually use it: pick your route

Most coverage stops at the announcement. The practical question is how you get your hands on it:

• Route 1 — hosted API (for almost everyone). One of the seven launch partners gives you a working model in minutes: an account, an API key, some usage credit. This is how freelancers and small teams should start.
• Route 2 — download the weights (for researchers and infrastructure teams). The full-precision checkpoint is roughly 1.6TB; the NVFP4 version needs around 530GB plus heavy GPU memory. Model cards cite at least 700GB of GPU memory for the quantized NVFP4 checkpoint — and at least 2TB for the full-precision BF16 one.
• Route 3 — two Mac Studios (for Mac-based labs). Only makes sense if you already live in the Apple ecosystem — and the 512GB Mac Studio wasn't widely available at launch, so this route is currently limited.

Ollama users: Ollama had early access and offers the model from day zero in partnership with Together AI, reporting good tool-calling results — so expect quantized versions in its library (an expectation, not a confirmed release) — but the full weights will never fit a typical machine.

What this means for readers (and what to watch)

If you're a developer or building with AI agents, Spark Open is worth your time: a 1M-context, multimodal, agent-focused model you can inspect, fine-tune and self-host without sending data to someone else's cloud. Writing effective prompts for agents becomes more valuable, not less, when the model is free to run.

If you're comparing it to the closed frontier options, see our ChatGPT vs Gemini vs Claude breakdown — and note that the open-weight lane has its own race: Mistral Large 4 is the European contender.

Three honest limitations:

1. Vendor benchmarks only. Independent evaluations will take weeks. Don't make infrastructure decisions on Meta's charts.
2. Hardware reality. "Runs on Mac Studios" means two very expensive, not-yet-shipping machines — this is not a laptop model.
3. License friction. The attribution and API-resale terms are deal-breakers for some companies that true open-source licenses wouldn't trouble.

The bottom line

Muse Spark Open is a real event: the US finally has a heavyweight open-weight answer to the Chinese models that have led this space all year, and the Apple partnership signals that frontier AI doesn't have to live in a data center. But it's a model for builders with infrastructure, not a toy for everyone — most readers should experience it through a (metered, per-token) API, not a download. The claims to watch are the independent benchmarks over the next few weeks; until then, Meta's "most capable" line is marketing, and the license is the fine print that decides whether "free" applies to you.

Sources & Further Reading

• Muse Spark Open model card — Hugging Face
• Muse Spark Open launch details — Together AI
• Muse Spark Open on Amazon SageMaker — AWS
• Meta launches Muse Spark Open, optimized for Mac Studio — 9to5Mac
• Meta Muse Spark Open Mac Studio details — MacRumors
• What Meta's open-weight license actually lets you do — dev.to