Meta Releases Muse Spark Open, an 816B Open-Weight Model

Meta released Muse Spark Open on October 15, 2026. The company calls it “the most capable US open-weight model to date.” It’s built for coding and agentic work. That means long-horizon tasks, complex software engineering, and multi-step reasoning.

The model is big. It has 816B total parameters but only 42B active. That’s a Mixture-of-Experts Transformer. It handles a 1M token context at inference. The vocabulary is 202,048 tokens. It’s natively multimodal on input—text, image, video, audio. Output is text only. It’s not an image or video generator.

Hardware requirements are steep

You need serious hardware to run this thing. The model card lists at least 700GB of accelerator memory for the NVFP4 version. The BF16 version needs 2TB. That’s not consumer gear.

There’s a reported setup that runs locally on two paired M5 Ultra Mac Studios via MLX. That gives a terabyte of unified memory. But there are caveats. Apple’s 512GB-memory configuration ships only in late October. Its price is unannounced. A 256GB M5 Ultra workstation with 1TB SSD costs $10,799. So comparable memory capacity today would cost tens of thousands of dollars. No measured generation speed or full setup cost has been published.

The supported serving stacks are broad. SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth. Hardware support includes AMD, Apple, Intel, and Nvidia. So it’s not locked to one vendor.

It’s not open source

The license is called the Muse Model License. It’s custom, not OSI open source. The key restriction: businesses whose group grosses over $20M in any 12 consecutive months must notify Meta and get a written agreement before offering the model as a paid model-serving service.

Internal deployment, research, and in-product use are exempt. So if you’re building a product that uses the model internally, you’re fine. If you’re selling API access to it, you need Meta’s sign-off. Products built on it must carry “Built with Muse Spark Open” attribution.

That’s a real difference from true open-source licenses. It gives Meta control over commercial reselling. For enterprises, it’s a US-developed downloadable option. For startups, it’s a potential legal hurdle.

Benchmarks: mixed results

Meta’s own reported table shows leads on some tests. It tops JobBench at 61.8. It leads MCP-Atlas at 89.2. On HLE-Diamond without tools, it scores 26.1. On visual ZeroBench, it gets 50.

But independent listings tell a different story. The Artificial Analysis Intelligence Index gives it a score of 39. That’s below MiMo-V2.6-Pro at 46, GLM-5.3 at 45, and Kimi K3 at 44. On terminal coding and extended software-engineering tests, Kimi K3 and GLM 5.3 score higher.

So it’s competitive but not the absolute top. It’s strong on agentic benchmarks and visual tasks. It lags on some coding and reasoning tests. These numbers come from Meta’s own reported table, independently unverified.

API access and pricing

You can access it via several providers. Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI all offer it.

Together AI’s serverless pricing: $1.25 per million input tokens. Cached input is $0.15 per million. Output is $4.25 per million tokens. That’s not cheap, but it’s in line with other large models.

There are context caps on some providers. GroqCloud reportedly caps context at 131,072 tokens. Fireworks offers the full 1,048,576. So if you need the long context, pick your provider carefully.

Weights download: a detailed form

Weights are available from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. The Hugging Face page is meta-models/Muse-Spark-Open-816B-A42B.

The access form asks for your full legal name, date of birth, and full organization name. That’s unusually detailed. But the same notice appears on Meta’s Llama 4 Scout page. So it’s a Meta pattern, not a one-off.

Why this matters now

Chinese open labs currently dominate open-weight downloads. Qwen, DeepSeek, and Moonshot AI’s Kimi are the big names. Hugging Face data says Chinese models took 41% of downloads in the year to March 2026. OpenRouter says open models handled 60% of US-origin token volume in September. Most of that came from Chinese labs.

Meta’s release gives US enterprises a downloadable alternative. It’s a US-developed model with a US-based company behind it. That matters for data sovereignty, compliance, and supply-chain concerns.

But the license is a sticking point. The $20M revenue threshold for paid serving is a real constraint. It’s not a fully open model. It’s a controlled release with attribution requirements.

Practical take

If you’re an enterprise looking for a US-sourced open-weight model for coding agents, Muse Spark Open is worth a look. The 1M context is a big deal for long-horizon tasks. The multimodal input is useful for real-world agentic work.

But check your hardware budget first. 700GB of accelerator memory is not a dev machine. The Mac Studio setup is clever but unproven. No one has published real generation speed or full setup cost yet.

And check the license. If you’re planning to resell model access, you need Meta’s written agreement. That’s a business decision, not a technical one.

For everyone else, the API route is easier. Together AI’s pricing is transparent. Fireworks gives you the full context window. Groq is cheaper but caps context.

Bottom line

Muse Spark Open is a serious release. It’s the biggest US open-weight model yet. It’s strong on agentic benchmarks and visual tasks. It’s not the top on all coding tests. It’s not fully open source. It’s expensive to run on your own hardware.

But it gives US enterprises a real option. That’s a shift in the open-weight landscape. Chinese labs have dominated downloads for a year. Meta is pushing back.

Whether that’s enough to change the download numbers is unclear. But it’s a concrete, downloadable alternative. That’s what matters.