Breaking News
Breaking: Meta Releases Muse Spark Open, Its 816B Open-Weight AI Model
Meta released Muse Spark Open on Oct 15 — an 816B open-weight AI model for coding, downloadable on Hugging Face, engineered to run on two Apple Mac Studios.
Meta just released Muse Spark Open, an 816-billion-parameter open-weight AI model it calls "the most capable U.S. open-weight model to date." Announced Thursday, October 15, 2026, from Meta's Superintelligence Labs unit, the model is built for coding and agentic work — and, in a notable twist, Meta engineered it to run on a pair of Apple Mac Studios sitting on a desk. The launch was covered by TechCrunch, MacRumors, and 9to5Mac, and the model card is live on Hugging Face.
What we know
- The specs. 816 billion total parameters (42 billion active per token via mixture-of-experts), a 1-million-token context window, and native multimodal input — text, image, video, and audio. Meta says it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
- Weights are downloadable. Unlike Meta's closed Muse Wonder models launched three days earlier, Spark Open's weights are free to download via Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Hosted inference is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
- It runs on Mac hardware. Working with Apple, Meta tuned the model for Apple's MLX framework. Two Mac Studios with M5 Ultra chips — each offering up to 512GB of unified memory — behave, per Meta, like "one supercomputer with a terabyte of unified memory." Meta demoed its Muse Code terminal, powered by Spark Open, shipping real changes to an open-source iOS app (bug fixes, new features, full build) and driving DaVinci Resolve end to end, reviewing footage frame by frame.
- The license has strings attached. Spark Open ships under Meta's own Muse Model License — open-weight, not open-source. Developers can use, modify, and sell products built on it for free, but must show a "Built with Muse Spark Open" credit, and companies selling model access through an API with more than $20 million in revenue need a separate deal with Meta.
- Meta's benchmark claims. In Meta's published scorecard, Spark Open leads rival open-weight models on 6 of 11 benchmarks, with Kimi K3 leading 4 — those are Meta's numbers, not an independent lab's. Meta says the model cleared its internal severe-risk safety thresholds.
- Meta's CEO thanked Apple's CEO. Mark Zuckerberg publicly thanked Apple CEO John Ternus, who responded that running a model of this scale "right from your desk" pushes "the boundaries of on-device AI."
What's unconfirmed
- Independent performance. Beyond Meta's own scorecard, full third-party evaluations are still early. Treat "most capable" as Meta's claim until more tests land.
- The hardware reality check. Meta's model card says the NVFP4 version needs at least 700GB of GPU memory (2TB for BF16). The 512GB Mac Studio configuration hasn't shipped yet — Apple says late October — and a 256GB M5 Ultra Mac Studio already costs over $10,000. This is not an everyday-computer model.
- How many will actually download it. The desktop milestone is real but narrow: the audience is developers and labs with serious hardware, not casual users.
Why it matters
This is a strategic about-face inside an about-face. Three days ago Meta went closed and paid with Muse Wonder; now it's back in the open-weights game at the high end — and framing the race explicitly against Chinese open-weight leaders like Kimi, GLM, and DeepSeek. For developers, it's a downloadable, tunable alternative to cloud-only frontier models with competitive coding abilities. For everyone else, the signal is bigger: the most powerful U.S. open model is meant to run on desktop hardware, not just in data centers. The honest version: the weights are real and downloadable today, the Mac Studio story is aspirational hardware, and the "most capable" crown still needs independent judges.
Sources: TechCrunch, MacRumors, 9to5Mac.