Muse Spark Open: Meta's Free 816B AI Model, Explained

On Thursday, October 15, 2026, Meta did something it hasn't done in over a year: it gave away the weights to a frontier-class AI model. Muse Spark Open is an 816-billion-parameter AI model you can download free, run on your own hardware (if your hardware happens to cost as much as a car), and build commercial products on. It arrived just three days after Meta's closed flagship, Muse Wonder — and it's already being called the strongest American open model in years.

But here's the thing nobody should gloss over: downloading this model is free. Running it is not even close to free. This article explains what Muse Spark Open actually is, what it can do, the two catches (hardware and license), and the three realistic ways you can use it — including the one most people will choose.

Quick answer: Muse Spark Open is Meta's 816B-parameter open-weight AI model (released Oct 15, 2026), built for coding and agentic work with a 1M-token context. The weights are free to download from Hugging Face under Meta's Muse Model License, but running them needs ~700GB of GPU memory minimum — so most people will use it through a hosted API at $1.25 per million input tokens on OpenRouter.

In this article:

What is Muse Spark Open?
What can it actually do?
How good is it? (benchmarks)
The two catches: hardware and license
How to actually use it (3 routes)
Muse Spark Open vs. the rivals
Verdict: who should care?
FAQs

What is Muse Spark Open?

Muse Spark Open is the open-weight version of Muse Spark — the model family behind Meta AI and its Muse Code agent. Meta's partner post on the PyTorch blog says it "shares the same base model as Muse Spark 1.3," with extra safety training so it holds up as a standalone checkpoint without Meta's own guardrails around it. It was released by Meta Superintelligence Labs, and Meta calls it "the most capable US open-weight model to date."

The specs, straight from the official model card:

816 billion total parameters, 42 billion active per token. It's a sparse mixture-of-experts model: 256 experts per layer, and a router picks 8 for each token. Think of it like a hospital with 256 specialists where only 8 see you per visit — the knowledge of all 816 billion parameters is available, but the compute cost stays near a 42B model.
1,048,576 tokens of context. That's roughly 2,500 pages of text in a single conversation — enough for an entire codebase or a very long legal document.
Natively multimodal input. It reads text, images, audio, and video, and writes text.
Reasoning that never sleeps. Reasoning is always on, with four effort levels — low, medium, high, and xhigh (the default). Meta recommends high or xhigh for complex problem-solving and agentic tasks.
Trained on 100+ languages, though the model card doesn't say which ones are strongly supported — worth testing if you work in a less-represented language.

It is not the same as Muse Wonder, the closed frontier model Meta launched three days earlier. Wonder stays API-only. Spark Open is the one you can download.

Independent researcher Nathan Lambert called it "the strongest American-made open model by quite some margin," adding that "Meta is the only American company with frontier open and closed models." Until this week, the open-weights leaderboard above it was entirely Chinese labs.

What can Muse Spark Open actually do?

Meta built this model for one job above all: agent work. Not chatting — doing. Reading a large codebase, planning a change, calling tools, checking the result, fixing its own mistakes. It can browse the web, run subagents, and use a computer across desktop, mobile, and browser.

The demos are the fun part. Running entirely on local hardware, a coding tool built on Spark Open shipped changes to Ice Cubes, an open-source Mastodon iOS app — and in another demo it drove DaVinci Resolve, examining individual video frames to pick highlights and produce vertical crops. That is a model editing a real iOS app and a real video project, on a desk, not a datacenter.

Here's the strange twist that has everyone talking: Meta worked with Apple to make it happen. These are two companies that spent years sniping at each other — the relationship soured badly after Apple's 2021 privacy changes hit Meta's ad business. Yet Meta optimized Spark Open for Apple's MLX framework so it can be split across linked Mac Studios. Mark Zuckerberg announced it on X with a photo of two Mac Studios on his desk, writing: "Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with @johnternus and Apple." Apple's CEO replied publicly, calling it a showcase for on-device AI.

For Apple, every demo proving serious AI runs on Mac hardware instead of a GPU datacenter is a win for its silicon story. For Meta, it's proof the model can genuinely run on a "desktop" — albeit a very expensive one. More on that in a moment.

How good is it? The benchmarks, honestly

Two things to know before the numbers: these are Meta's own published benchmark scores, not independent verification. And the picture is mixed — which is actually more interesting than a clean sweep.

Tool use: excellent. On MCP-Atlas (a tool-use benchmark) it scored 89.2, beating Kimi K3 (82.3), GLM 5.3 (84.2), and Qwen3.8 (84.5). On JobBench it led again at 61.8. This fits Meta's "built for agents" story.

Pure coding: behind the Chinese open models. On Terminal-Bench 2.1 it scored 76.4 against Kimi K3's 85.0; on DeepSWE, 60 against Kimi's 68. So if your main job is writing code in a terminal, Kimi K3 still scores higher — on Meta's own chart.

Hallucinations: its best trick. On the AA Omniscience test, which penalizes made-up answers, it scored 25 — top of its class. And on safety, the gap is dramatic: on the StrongREJECT jailbreak test, Meta reports 0.5 for Spark Open against 62.9 to 66.6 for its rivals, with 99.2% of jailbreak attempts refused in Meta's cyber tests. That makes sense for a company releasing weights anyone can download — but again, remember these are Meta's tests of its own model.

The independent Artificial Analysis index scored it 38.9 (rounded to 39) — the highest score for a US lab's open-weight model, but 10th overall, behind nine models from Chinese labs. It lands level with DeepSeek V4.1 Flash, and nine points below its closed sibling Muse Spark 1.3.

Bottom line: a strong agent model that picks the right tool and invents fewer things — not the best open model for pure coding.

The two catches: hardware and license

Catch 1: the hardware

This is the big one. Only 42B parameters work on each token, which helps speed — but all 816B have to sit in memory. The sparse design helps inference cost, not RAM.

What you actually need, per Meta and its partners:

Full BF16 version (~1.65 TB download): about 2 TB of VRAM — think eight NVIDIA B300 GPUs.
Compressed NVFP4 version (~537 GB download): at least 700 GB of VRAM — a 4-GPU Blackwell node, or eight H200s.
Apple MLX version: two Mac Studio M5 Ultras with 512GB unified memory each, linked over Thunderbolt 5. The 512GB Mac Studio hasn't even shipped yet (expected later this month), and the existing 256GB version costs over $10,000.
Smallest community version (Unsloth GGUF, ~2-bit): about 300 GB of RAM+VRAM combined via llama.cpp.

So no — it will not run on a gaming GPU. A PC with an 8GB graphics card is roughly a quarter of the way to the smallest version. Realistic home ceiling is models around 27–30B parameters. Meta's own answer for local use is a different model entirely: Muse Glimmer, a 30B model released in August under Apache 2.0, designed for a single consumer GPU.

Catch 2: the license

"Downloadable" is not the same as "open source." Muse Spark Open ships under Meta's own Muse Model License, plus an Acceptable Use Policy. The terms are more generous than many expected, with conditions that matter:

Attribution: you must show "Built with Muse" in your product's UI, docs, website, or settings.
Big API businesses: if you sell API access to the model and your group's gross revenue exceeds $20 million in any 12 months, you must notify Meta and sign a separate agreement first. Internal use, research, and building it into your own product are exempt — so for almost every small developer, this never applies.
Gated download: on Hugging Face you must provide your full legal name, date of birth, and organization — and you can't edit the form afterward. An unusual hoop, but it is what it is.
Fine-tuning, distilling, merging, and selling things built with it are all allowed; Meta claims no ownership of what the model writes.

Researcher Nathan Lambert called the custom license "an unnecessary strategic risk" — a fair critique of a moment when Meta could have gone fully permissive.

How to actually use it: three routes

Route 1: Download the weights (free, hard). On Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Free to download — then bring your 700GB+ of GPU memory.

Route 2: Run it on linked Mac Studios (expensive). Meta's research blog calls it the first model in its class to run on a desktop — two M5 Ultra Mac Studios behaving as one supercomputer with a terabyte of unified memory. Impressive, but two unreleased $10,000+ machines make this a press-demo setup, not a practical one, for most people.

Route 3: Use a hosted API (the practical one). Day-zero launch partners include Together AI, Baseten, Fireworks AI, Groq (limited preview), Modal, Nebius, and OpenRouter. On OpenRouter it's priced at $1.25 per million input tokens and $4.25 per million output tokens — the same price as the closed Muse Spark 1.3, and roughly a third of the input price and a fifth of the output price of Muse Wonder ($4 input / $20 output). For most readers, this is the only route that makes sense: no hardware, pay-as-you-go.

Meta's own pages don't publish partner prices, so compare providers before committing. And with models arriving almost daily now, pay-as-you-go beats any commitment.

Muse Spark Open vs. the rivals

Model	How you get it	Key specs	Price signal
Muse Spark Open (Meta)	Downloadable weights + APIs	816B total / 42B active, 1M context, multimodal	Free to download; $1.25/$4.25 per 1M tokens via API
Muse Wonder (Meta)	API only	Closed frontier model, Meta's strongest system	$4/$20 per 1M tokens
Kimi K3 (Moonshot)	Open weights	Leads Meta's chart on pure coding benchmarks	Cheaper open alternative for coders
Gemini 4 Argon (Google)	API only	Reached general availability this week	$2/$10 per 1M tokens until Nov 30, then doubles

Read our Gemini 4 Argon breakdown for Google's counter-move this same week, and our GPT-6 with Intelligent UI explainer for OpenAI's side of the October model race.

Verdict: who should care about Muse Spark Open?

Care about it if: you build AI agents, you want a strong model you can host in your own cloud account (sovereignty, compliance, or cost reasons), or you care about safety and fewer hallucinations — its tool-use scores and jailbreak resistance are genuinely class-leading.

Skip it if: you want the absolute best coding model (Kimi K3 leads there), or you imagined running an 816B model on a home PC. The honest summary: "free to download" and "runnable by you" are two very different things here.

The bigger story isn't the specs — it's the direction. This is Meta's first major open release since Llama 4, and Meta is now the only American company fielding both frontier-class closed and open models. In an open-weights race where Chinese labs had been setting the pace unchallenged for 18 months, that's the headline that actually matters.

Frequently asked questions

Is Muse Spark Open really free?

The weights are free to download from Hugging Face under the Muse Model License, and you can fine-tune and commercialize products built on it (with "Built with Muse" attribution). But running it yourself needs hundreds of GB of GPU memory, so the practical cost is the API route — $1.25 per million input tokens on OpenRouter.

Can I run Muse Spark Open on my laptop?

No. The smallest community version needs about 300GB of combined RAM and VRAM; the full version needs around 2TB of GPU memory. A normal laptop or gaming PC is far short. Meta's Muse Glimmer (30B, Apache 2.0) is the Meta model actually aimed at local use.

What's the difference between Muse Spark Open and Muse Wonder?

They launched three days apart in October 2026. Muse Wonder is Meta's closed frontier model — API-only, more capable overall, priced at $4/$20 per million tokens. Muse Spark Open is the downloadable open-weight sibling, built for coding and agentic work.

Is Muse Spark Open actually open source?

Not in the strict sense. It's "open weights" under Meta's custom Muse Model License: free to download, fine-tune, and commercialize, but with a "Built with Muse" attribution requirement, an acceptable use policy, and a separate agreement for API businesses over $20M in revenue. There is no OSI-approved open-source license here.

How good is Muse Spark Open at coding?

Good, not best-in-class. Meta's own benchmark tables show it leading on tool use (MCP-Atlas 89.2) and hallucination resistance, but trailing Kimi K3 and GLM 5.3 on pure coding benchmarks like Terminal-Bench 2.1. It's an agent model first, a coding model second.

Where can I try Muse Spark Open without a server?

Through hosted APIs from day-zero partners: Together AI, Baseten, Fireworks AI, Groq (limited preview), Modal, Nebius, and OpenRouter. OpenRouter lists it at $1.25 input / $4.25 output per million tokens with the full 1M context.

Methodology

This article was researched on October 18, 2026, from Meta's official Muse Spark Open model card and developer documentation, Meta Superintelligence Labs' research blog (Oct 15, 2026), the PyTorch Foundation's partner announcement, Artificial Analysis' Intelligence Index, and reporting from independent outlets covering the release. All benchmark figures are the vendors' own published numbers and have not been independently verified; API prices were checked via OpenRouter listings and may change. For a deeper spec-by-spec walkthrough, see Michael Leung's breakdown of the model card, benchmarks, and hardware and this developer guide to running Muse Spark Open. This site is not sponsored by or affiliated with Meta. If you found this useful, see more in AI News and AI Software, or read about the best free AI tools for students. The AI Tools · About us · Contact

Search This Blog

The AI Tools