Meta Released Its Model's Weights for Anyone to Download. Running It Takes Two Mac Studios. (2026)

Meta put the weights for its newest AI model online on October 15, 2026, and called it "the most capable US open-weight model to date" 1. Anyone can download the file and run it on their own hardware. Meta says two Apple desktops are enough to hold it 2.

What the model is

Muse Spark Open is a natively multimodal model with 816 billion total parameters and 42 billion active on each token, and a context window of one million tokens 1. Text, images, video and audio go in; text comes out. The weights download from Hugging Face, Amazon SageMaker AI, Google Cloud's model garden and Microsoft Foundry, and seven hosted platforms — Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI — serve it from day one 1.

Meta's own demonstrations pair the model with Muse Code, its coding terminal: it shipped bug fixes, new features and a full build to Ice Cubes, an open-source Mastodon client, and drove DaVinci Resolve end to end, reviewing footage frame by frame and cropping it vertically 1.

Running it on a desk

The desktop configuration is two Mac Studios with Apple's M5 Ultra chip, running Apple's MLX framework. Apple's unified memory gives each machine up to 512GB, which is how the pair holds an entire model: about a terabyte of memory between them, at 1.2 terabytes per second of bandwidth 1.

That is also the catch. Running it locally takes on the order of a terabyte of memory, and the 512GB Mac Studio does not ship until late October, with no announced price 3. Apple lists a 256GB M5 Ultra workstation with a 1TB SSD at $10,799; four of those would provide the same terabyte for about $43,196 before tax and accessories 3. The model card asks for at least 700GB of accelerator memory in the NVFP4 format and 2TB in BF16 3.

Apple and Meta, on the same page

The two companies are normally rivals and here backed each other publicly. Mark Zuckerberg posted that Muse Spark Open "runs great on new Mac Studios thanks to work with @johnternus and Apple", and Apple's chief executive John Ternus replied to congratulate the team 4. Meta's chief AI officer, Alexandr Wang, told Axios the companies had been working together for several months, and that he was glad to see them getting along. Apple also published a technical post of its own about MLX 4.

The fine print

The weights come under the Muse Model License, which is free for most developers 1. It allows running, modifying, fine-tuning, distilling and redistributing the model, with conditions: once a business and its affiliates pass $20 million in aggregate gross revenue over twelve consecutive months, serving the model or its derivatives requires notifying Meta and a separate written agreement. Products must carry the attribution "Built with Muse Spark Open" 3.

Hugging Face's download gate asks for a full legal name, a date of birth and the organization's corporate identifiers, and warns that failing to follow the instructions may affect access to other models on the site 5. Hosted access varies by provider: Together AI lists $1.25 per million input tokens and $4.25 per million output tokens, while Groq caps the context at 131,072 tokens against the model's advertised million 3.

The scoreboard

On the benchmark table Meta published, Spark Open leads JobBench at 61.8 and MCP-Atlas at 89.2, and sits behind GLM 5.3, Kimi K3 and Qwen3.8-2.4T-A95B on terminal coding, extended software-engineering work and GDPVal-AA 3. Meta's card reports an Artificial Analysis Intelligence Index score of 39, while independent listings put six Chinese models at 40 to 46 — a comparison VentureBeat notes stays provisional until Meta's released checkpoint is independently assessed 3.

The same day

Nvidia shipped the open-weight Nemotron 3.5 Super VL, a 120-billion-parameter model built for document, chart and image understanding, plus a Bolt family that puts a 74-billion-parameter model into 24GB of memory 6. The models use its OpenMDW licence, which permits commercial use, fine-tuning and redistribution, and Nvidia quotes $4,500 per GPU per year for production software support 7.

Microsoft said it finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active, and that it outperforms every open-weight model Microsoft tested 8. That model is not being given away: it is still in post-training, and Microsoft says it will license the weights to trusted partners through Foundry 8.

"Open Weights" is an original rap song and music video built from those facts — Southern crunk, one song, one day's news.

Visual material: the desk, the data centre aisle, the silicon, the server rack, the contract and the engineers are free-stock photographs and video under the Pexels License, which requires no credit; the release pages from Meta AI Research, NVIDIA and Microsoft AI and the Hugging Face model card are screenshots of the pages linked above; the photograph of the two Mac Studios is from Mark Zuckerberg's post on X; everything else on screen is drawn for this episode.