Meta Muse Spark Open: Specs, Benchmarks, and How It Compares
Meta's 816B Muse Spark Open claims the top US open-weight model spot. Here's how it benchmarks, licenses, and runs against Chinese rivals.
What is Meta Muse Spark Open?
Muse Spark Open is a mixture-of-experts language model Meta released as open weights, with 816 billion total parameters and 42 billion active per token. It’s natively multimodal (text, image, video, and audio input, text output), supports a 1 million token context window, and is listed on Hugging Face under the model ID meta-models/Muse-Spark-Open-816B-A42B. Meta calls it the most capable US open-weight model available, a claim that holds up against other American releases but not against the current crop of Chinese open models.
TL;DR
- Muse Spark Open is an 816B-parameter mixture-of-experts model with 42B active parameters, a 1 million token context window, and native multimodal input across text, images, video, and audio.
- On the Artificial Analysis Intelligence Index, it scores 39, ahead of other US open releases but behind Chinese open models like Mimo 2.6 Pro (46), GLM 5.3 (45), Kimi K3 (44), and Qwen 3.8 (40).
- It beats Reflection’s Beam model (30 on the same index, 501B total/23B active parameters) on most published benchmarks, though Beam is smaller, more token-efficient, and ships under a cleaner Apache 2.0 license.
- Muse Spark Open uses Meta’s own custom license, which requires a “built with Muse Spark Open” attribution and a separate commercial agreement for model-as-a-service businesses earning over $20 million a year.
- The model is not what powers Meta’s Muse personal-agent app; that app runs a customized, fine-tuned version of the separate closed Muse Spark 1.3 model.
- Meta also shipped two closed frontier models, Muse Wonder and Muse Wonderful, in the same stretch, scoring 52 and 56 on the Intelligence Index, within striking distance of Claude Opus 5.5’s 58.
- Running Spark Open locally needs serious hardware (roughly 530GB for an NVFP4 checkpoint, or around 300GB with a community 2-bit quantization); most people will access it through hosted providers on Open Router instead.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
How does Muse Spark Open benchmark against other open models?
Spark Open sits in a specific spot: it’s the strongest US open-weight model, but it is not the strongest open-weight model overall. On the Artificial Analysis Intelligence Index, it scores 39. Every model ahead of it on that index right now is Chinese: Mimo 2.6 Pro leads at 46, followed by GLM 5.3 at 45, Kimi K3 at 44, and Qwen 3.8 just edges past Spark Open at 40.
That gap is worth sitting with. Meta’s own marketing language is careful to say “most capable US open-weight model,” not “most capable open-weight model,” and the benchmark numbers explain why. The frontier of open-weight AI right now is being set by Chinese labs, and Meta’s release narrows the gap for US-built models without closing it.
Against other American open releases, though, Spark Open clearly wins. Reflection’s Beam model, which also published its weights recently, scores 30 on the same index. Looking at task-specific benchmarks, Spark Open beats Beam on tool use (MCP Atlas: 89 vs. 79) and on DeepSeek-style evaluations (60 vs. 44). Beam does edge ahead on terminal bench, 80 to 76.
That said, the comparison isn’t perfectly apples-to-apples. Beam is a much smaller model, 501 billion total parameters with 23 billion active, and Reflection built it specifically to be token-efficient. A smaller model losing on raw benchmark scores while winning on efficiency and cost is a normal tradeoff, not a flaw.
What’s the licensing difference between Spark Open and Beam?
This is where the two models diverge in a way that matters for anyone building a product. Beam ships under Apache 2.0, a fully permissive open-source license with essentially no strings attached.
Spark Open ships under Meta’s own Muse license. For most individual developers and small projects, the terms won’t change much day to day. But there are two conditions worth knowing:
- You need to display “built with Muse Spark Open” somewhere in your product.
- If you’re operating a model-as-a-service business generating more than $20 million a year in revenue, you need a separate commercial agreement with Meta before you can offer the model that way.
So the practical framing is: Spark Open is the more capable model but comes with a heavier, Meta-specific license and a much bigger footprint. Beam is smaller, cheaper to run, and cleanly licensed. Which one makes sense depends on whether raw capability or deployment simplicity matters more for the use case.
How do you actually run Muse Spark Open?
At 816 billion total parameters, this isn’t a model most people run on a laptop. Meta worked with Apple to validate a high-end setup: two Mac Studios with M5 Ultra chips and 512GB of memory each, which Apple reportedly measured at around 25 tokens per second. That’s a specialized, expensive configuration, and new M5 Ultra hardware has had long lead times, so this path isn’t realistic for most developers in the near term.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
For people not buying dedicated Mac hardware, Meta has released an NVFP4 quantized checkpoint weighing in around 530GB. The community has gone further: Unsloth released a 2-bit quantized version that shrinks the footprint to roughly 300GB. Both are still substantial, requiring serious multi-GPU setups or high-memory unified architectures.
For most developers, the realistic path is a hosted inference provider. Spark Open went up on Open Router on day one, and multiple providers including Together AI, Fireworks, Base Ten, Nebius, and Modal are already serving it, all at the same published rate of $1.25 per million input tokens and $4.25 per million output tokens. Notably, Meta itself isn’t one of the providers hosting Spark Open on Open Router, unlike with its closed models.
Is Muse Spark Open the same model powering Meta’s Muse app?
No, and this is a common point of confusion. Muse Spark Open is a distinct release from whatever powers Meta’s Muse personal-agent app, which has become one of the most-downloaded AI apps in the US since launching on September 8th. The app reportedly passed 2.5 million downloads within a couple of weeks and reached the top of the Apple App Store, each user getting an isolated virtual machine for their agent and data.
The Muse app actually runs on a customized, fine-tuned version of the separate Muse Spark 1.3 model, not Spark Open. According to comments from Meta Superintelligence Labs’ Alexander Wang, the agent was built over roughly seven months by a team of under 200 people, using a method of listing over 100 specific required behaviors, building evaluation tests for each one, and refusing to ship until the model passed every single eval at a “launch blocking” threshold. That approach, treating agent reliability as a checklist of testable behaviors rather than a single benchmark score, is a useful blueprint for anyone building agentic products, independent of which underlying model is used.
How does Spark Open compare to Meta’s closed models?
Spark Open isn’t close to Meta’s best work. In the same window as the open release, Meta also launched two closed frontier models, Muse Wonder and Muse Wonderful, benchmarked against GPT-6 Astra and Claude Opus 5.5. On the Artificial Analysis Intelligence Index, Wonder scores 52 and Wonderful scores 56, just two points behind Opus 5.5’s 58. Sitting below those is the closed Muse Spark 1.3 at 48, nine points above the open Spark release.
Access to these closed models is uneven. Muse Spark 1.3 is available on Meta’s API and Open Router at $1.25 input / $4.25 output, with a steeply discounted $0.10/$0.20 tier available if you allow Meta to train on your prompts and outputs. Muse Wonder is priced at $4 input / $20 output, matching Opus 5.5 exactly. Muse Wonderful isn’t available via API at all, it’s locked inside Meta’s Muse Code product on $50 or $200 monthly plans, which limits independent testing of Meta’s strongest performance claims. On LM Arena’s agent-specific leaderboard, Wonder reportedly ranked 12th despite placing second on general text leaderboards, suggesting Meta’s frontier strength doesn’t fully carry over to agentic tasks yet.
Frequently Asked Questions
What does “816B-A42B” mean for Muse Spark Open?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
It describes the model’s mixture-of-experts architecture: 816 billion total parameters across all experts, with only 42 billion active for any given token. This keeps inference cheaper than a dense model of the same total size, since only a fraction of parameters are used per forward pass.
Is Muse Spark Open free to use commercially?
It’s free to download and use under Meta’s Muse license, which requires attribution (“built with Muse Spark Open”) and, for large-scale model-as-a-service businesses earning more than $20 million annually, a separate licensing agreement with Meta.
Can I run Muse Spark Open on consumer hardware?
Not realistically. Even quantized, the model requires hundreds of gigabytes of memory (around 530GB for the NVFP4 checkpoint or about 300GB for a community 2-bit version). Most developers will access it through hosted inference providers rather than running it locally.
Is Muse Spark Open better than Chinese open models like Qwen or GLM?
No. On the Artificial Analysis Intelligence Index, it scores 39, behind Mimo 2.6 Pro (46), GLM 5.3 (45), Kimi K3 (44), and Qwen 3.8 (40). It leads among US open-weight models specifically, not open models overall.
Does Muse Spark Open power Meta’s Muse app?
No. The Muse app runs a separately fine-tuned version of Meta’s closed Muse Spark 1.3 model, built specifically for agentic tasks like email, bookings, and form-filling. It is not the same model as the openly released Spark Open.
