Meta Releases Muse Spark Open: 816B Open-Weight AI Model Specs & Download

Meta just did what it hasn’t done at this scale in a long time: it released the weights. Muse Spark Open, launched October 15, 2026 by Meta Superintelligence Labs, is the open-weight version of the Muse Spark model that had been proprietary since its April debut. It’s an 816-billion-parameter mixture-of-experts model aimed at coding and long-running agent work — and Meta calls it the most capable open-weight model yet released by a US lab.

The headline number is big, but the real story is where it runs: Meta demoed the model serving on two Apple Mac Studios, pitching an 816B-parameter model as something that can live on a desk rather than in a data center. Here’s the specs, the benchmarks, the license fine print, and how to get it.

Muse Spark Open Specs: 816B Parameters, 42B Active

The model is a sparse mixture-of-experts (MoE) transformer. The key numbers, per Meta and the Hugging Face model card:

• Total parameters: 816 billion; active per token: about 42 billion (8 of 256 experts fire per token in each MoE layer)
• Architecture: 62 layers, 8,192 hidden dimension, grouped-query attention (128 query heads, 16 KV heads)
• Context window: up to 1 million tokens — enough to feed in a large repository, a lengthy tool-call history or extended multimodal input in one prompt
• Modalities: natively multimodal — takes text, images, video and audio as input, returns text (a separate 1.8B-parameter Perception Encoder handles visual input)
• Training data: covers more than 100 languages; 202,048-token vocabulary
• Reasoning control: a reasoning_effort setting (low, medium, high, xhigh — default xhigh)

The starting point was Muse Spark 1.2, shipped in August, and Meta added extra safety post-training so the open model holds up without the guardrails of Meta’s own services. Meta says the release cleared the release thresholds in its Meta Superintelligence Scaling Framework, with a full evaluation report still to come.

Muse Spark Open Benchmarks: Where It Leads and Lags

Independent evaluations haven’t verified Meta’s claims yet, but the launch benchmarks paint a split picture:

Where it leads:

• MCP-Atlas: 89.2 — the top score in Meta’s comparison table (agentic tool use)
• JobBench: 61.8 — more than twice Inkling’s 31.1 (the US open model from Thinking Machines Lab) and NVIDIA’s Nemotron 3 Ultra at 27.2

Where it lags:

• Terminal-Bench 2.1: 76.4 — behind Chinese open models Kimi K3 (85), GLM 5.3 (83.9) and Qwen3.8 (82)
• Artificial Analysis Intelligence Index: 39 in Meta’s own table — below GLM 5.3 (45) and Kimi K3 (44)

The honest read: it looks genuinely strong for agentic tool use and long-horizon coding workflows — exactly what Meta built it for — while pure code-generation benchmarks still favor the Chinese open models. Treat Meta’s “most capable US open-weight model” line as the publisher’s own claim until independent testing lands.

The Muse Model License: What the Fine Print Says

This is an open-weight release, not open source — and the license matters. The release ships under Meta’s Muse Model License, which:

• Allows use, modification, fine-tuning and distribution — and Meta says many developers can use it for free
• Requires products built on it to display “Built with Muse Spark Open”
• Requires a separate agreement with Meta for Model-as-a-Service offerings once the provider’s aggregate gross revenue exceeds $20 million in any 12 months (internal use and research are excluded)

Developers should read the license terms before treating the release as unrestricted open-source software. It’s more permissive than a pure API deal — you get the weights — but it’s Meta’s license, not MIT.

How to Download It: Hugging Face, Cloud and APIs

Meta is distributing the weights and access widely:

• Weights: Hugging Face, plus cloud model catalogues — Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models
• Checkpoint sizes: roughly 822 GB for the BF16 build, about 484–493 GB for the NVFP4 version
• Hosted API access: seven platforms including Together AI, Groq, OpenRouter, Baseten, Fireworks AI, Modal and Nebius
• Inference frameworks: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth; Meta points fine-tuners at PyTorch’s TorchTitan

The practical upshot: most developers will call it through an API, not download 800 GB of weights. The self-hosting story is aimed at teams that want to control their infrastructure and data.

Can You Run It on a Mac?

This was Meta’s showpiece demo: the model running on two Apple Mac Studio desktops with M5 Ultra chips, linked into one system by Apple’s MLX framework. The memory math, per Apple’s Machine Learning Research post:

• BF16 precision: more than 1.6 TB just for parameters — can’t fit on any Mac
• NVFP4 (4-bit): about 493 GB — with working memory, Apple says it needs two or more Mac Studios (512 GB unified memory each, ~1 TB pooled)
• Speed: Apple’s table lists 15.74 tokens/second on two machines (pipeline parallelization), 24.99 with tensor parallelization, and 36.73 on four machines, at an 8,192-token context
• Lighter option: the Unsloth 2-bit GGUF build trims setups to about 300 GB

Meta calls it the first model in its class to run on a desktop. The honest version: it runs on two $4,000+ desktops wired together — a desk, yes, but not most people’s desk. If you’re exploring open AI tools and want something that runs on ordinary hardware, see what’s new in ChatGPT’s GPT-6 Intelligent UI — or check out this week’s other big AI launch, Amazon’s Bee Frontier, which puts AI on your wrist instead of your server rack. And if you’re shopping for a camera flagship rather than an AI model, our OPPO Find X10 preview covers the October 21 global launch.

Should You Use It?

The open-weight release is built for one audience: developers building coding assistants and agentic tools who want to download the weights, control their infrastructure, or inspect behavior directly — plus researchers who need an open model that handles long, multimodal, multi-step work. If that sounds like you, the hosted API route (Groq, Together AI, OpenRouter) is the fastest way to try it.

If you’re not shipping code or agents, the Mac Studio demo is the more fun story: an 816B-parameter model that runs on a desk is genuinely new territory for open weights, even if that desk needs two machines and a terabyte of memory.

Bottom line: 816B total / 42B active parameters, 1M-token context, native multimodality, weights on Hugging Face — Meta’s biggest open-weight release to date, strong on agent benchmarks, with a license that asks you to read the fine print.