Meta releases 816B-parameter Muse Spark Open model optimized for Macs
13 sources · The Decoder · The Verge AI · Hacker News front page (AI)- Boom: Meta releases Muse Spark Open, an 816B-parameter multimodal model, on October 15, 2026
- Boom: Meta and Apple collaborated on optimizing the model for M5 Ultra Mac Studios using MLX
- Doom: Running the model locally requires 1TB of RAM across two Mac Studio units
- Boom: Model handles text, images, video, and audio with up to one million tokens of context
- Neutral: Meta benchmarks Muse Spark Open against Chinese open-weight models Qwen, GLM, and Kimi
The story in full
Meta released Muse Spark Open on October 15, 2026, an open-weight multimodal model with 816 billion parameters. The company describes it as the "most capable US open-weight model to date." It processes text, images, video, and audio with up to one million tokens of context, and requires 1TB of RAM to run locally. Meta and Apple collaborated on the launch, with the model optimized to run on two Apple M5 Ultra Mac Studios using the MLX framework.
Meta positioned Muse Spark Open as a direct competitor to open-weight Chinese models including Qwen, GLM, and Kimi, saying benchmark performance is roughly on par with those systems. The collaboration between Meta and Apple is notable given the two companies are typically considered rivals. The model was also welcomed onto Hugging Face, where it is listed as an open-weight frontier multimodal mixture-of-experts model.
Analysis
393 wordsOn October 15, 2026, Meta released Muse Spark Open, an 816-billion-parameter open-weight multimodal model capable of processing text, images, video, and audio within a context window of up to one million tokens. The launch involved a notable collaboration with Apple, with the model optimized to run locally on two Apple M5 Ultra Mac Studios using the MLX framework. Meta describes it as the most capable US open-weight model to date and benchmarks it against Chinese open-weight systems including Qwen, GLM, and Kimi, claiming roughly comparable performance. The model is listed on Hugging Face as a multimodal mixture-of-experts architecture.
The Meta-Apple pairing is the detail most outlets have highlighted, given the two companies have long been considered rivals in consumer technology. Beyond the corporate dynamics, the release raises a genuine question about where frontier AI inference is heading: toward centralized cloud infrastructure or toward powerful local hardware. Running Muse Spark Open locally requires 1TB of RAM split across two Mac Studio units, a setup costing well over ten thousand dollars, which frames the local-run proposition as accessible to a narrow audience at best.
Pro-AI voices are enthusiastic. Accounts like Techmeme X Chatter flagged the 816-billion-parameter count and the mixture-of-experts architecture as reasons to expect real speed advantages, and Mazdak argued that running agents on local hardware rather than renting cloud compute is the direction the field is moving. Hermes at Goke framed this as a shift from training scale to edge deployability. Critics are less impressed. Public Citizen warned that the model still creates a data file accessible to Meta, directly undercutting the privacy case for local deployment, and incognitoastronaut questioned why Apple would partner with a company associated with repeated data controversies. The Techmeme X Chatter account also noted that the Muse Spark Open license is not fully open, a point the middle-ground camp echoed. The New Stack confirmed the model matches Chinese rivals on agentic tasks but flagged the licensing issue, and Space Captain Lispegistus cited a high refusal rate on certain queries as a reason for skepticism about real-world utility.
The clearest open question is whether the license terms will limit adoption by inference providers and developers who need commercial flexibility. How Meta addresses those terms, and whether independent benchmarkers replicate the claimed parity with Qwen and its peers, will determine how seriously the research and enterprise communities take this release.
What Pro-AI voices are sayingMeta's Muse Spark Open is celebrated as a landmark open-weight release, praised for local privacy, commercial freedom, and its Apple collaboration enabling Mac-native deployment without cloud or Nvidia dependence. Some see it as the start of a shift toward edge-deployable agentic AI.
Quote 1 of 13What Anti-AI voices are sayingCritics distrust Meta as a partner for Apple given its data scandal history, and worry about the restrictive license and the steep hardware cost of two machines exceeding $10,000. Some dismiss the release as another costly all-or-nothing gamble.
Quote 1 of 11What Middle Ground voices are sayingThe model leads US open-weight rivals and matches Chinese models on agentic tasks, but its license is not fully open and it lags some competitors on coding benchmarks. The Meta-Apple deal is seen as mutually beneficial, though spatial reasoning remains a published limitation.
Quote 1 of 10Add your take
0 reader votesSign in with Google to pick a side and post. Your vote moves the story's Doom / Boom score.
More Pro-AI reactions (12)
“Meta releases Muse Spark Open, a new SOTA open model.”
Unsloth AI, Bluesky · 17:39 UTC“The partnership shows how far the AI race is reshaping even the most adversarial relationships in tech.”
Yahoo Finance, Bluesky · 17:53 UTC“Meta just keeps winning. Ended up teaming up with Apple too!!! Beginning of the Golden Age of Local Superintelligence”
Techmeme X Chatter, Bluesky · 17:38 UTC“Meta is so back! They are finally releasing Meta Muse Spark open weight. Best US open weight model 816B total / 42B active parameters 1M token context window Natively multimodal across text, image, video, audio Love to see!”
Techmeme X Chatter, Bluesky · 17:56 UTC“Very happy to see Muse Spark Open!”
Eric Florenzano, Bluesky · 17:46 UTC“Meta's 816B param Muse Spark is open wieghts 🤯 1M-token context window. And MoE, so likely relatively fast. Now I'm really curious what my cluster can do!”
Techmeme X Chatter, Bluesky · 18:03 UTC“Meta is open sourcing the weights for Muse Spark Open, the most capable US open-weights model to date.”
PyTorch, Bluesky · 17:57 UTC“running agents on your own desk instead of renting cloud is where this is heading.”
Mazdak, Bluesky · 18:06 UTC“The agentic race is shifting from training scale to edge deployability.”
Hermes @ Goke, Bluesky · 18:03 UTC“Meta ships Muse Spark Open, its answer to China’s best open-weight models”
davidi, Bluesky · 18:30 UTC“Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.”
Techmeme X Chatter, Bluesky · 17:12 UTC“Meta released a model that runs better on apple silicon than apple.”
Techmeme X Chatter, Bluesky · 18:08 UTC
More Anti-AI reactions (10)
“It's a repeat of the Metaverse gamble, just instead of not being able to code basic VR functionality, he can't code a basic chatbot that won't turn creepy.”
SquidRadio, Bluesky, skeptic · 04:42 UTC“love making trillion dollar all-or-nothing gambles like im a plucky startup instead of a core tech sector company”
Sagittarius A*, Bluesky · 04:49 UTC“I'm sure it'll all be fine, after all these are the geniuses who pissed away 80 billion on the metaverse, so they know what they're doing.”
Peter Gordon 🇺🇦, Bluesky, skeptic · 05:59 UTC“Automating the workforce may be the reason why uptake is low too: The majority of us have no employees.”
NaffNaffBobFace, Bluesky · 05:17 UTC“You would think Apple would know better considering every day is a new data scandal from Meta”
incognitoastronaut.bsky.social, Bluesky · 17:26 UTC“Meta and Muse can go die in a fire for all I care.”
madbodger.bsky.social, Bluesky · 17:26 UTC“if the "muse spark open" doesn't make it behave like someone who thinks they are incapable of being wrong no matter how smart it is it will be a useless model”
ayla, Bluesky · 18:25 UTC“If this is supposed to make it less weird, it's not working.”
The Cleric, Bluesky · 13:03 UTC“Over $10,000 in not one, but two computers”
Soulfire The Wolf, Bluesky · 17:54 UTC“The Meta Muse Open License is trash. Thankfully no one really cares about licenses, expect the inference providers who will happily comply.”
Techmeme X Chatter, Bluesky · 17:48 UTC
More Middle Ground reactions (9)
“it's substantial refusal rate on "dangerous questions" I'm not convinced this is worth my attention.”
Space Captain Lispegistus 🇪🇺🇧🇬🏴☠️, Bluesky, skeptic · 18:14 UTC“just popping downstairs with a few friends to go ice skating in hell.”
John Philpin, Bluesky · 17:59 UTC“Meta’s Muse Spark Open is an 816B open-weight model that beats US rivals and matches Chinese models on agentic tasks, but its license isn’t fully open.”
The New Stack, Bluesky · 17:30 UTC“It's also a mutually beneficial deal. Meta gets more researchers and developers to use its AI model and Apple gets to sell more Macs.”
Techmeme X Chatter, Bluesky · 18:38 UTC“Is the future of enterprise AI a desktop, not a data centre?”
Ryan Daws, Bluesky · 17:46 UTC“Meta lists 3D spatial reasoning as a known limitation in the Muse Wonder model card, so the benchmark-topping numbers come with a boundary the company itself drew and published.”
yomimonoid.bsky.social, Bluesky · 22:56 UTC“보안과 비용 면에서 유리합니다.”
푸른곰, Bluesky · 17:07 UTC“Muse Wonder Full: not as smart, but better at coding tasks, doesn't have a place in my toolkit.”
Sherveen, Bluesky · 13:42 UTC“Meta udostępniła wagi multimodalnego modelu Muse Spark Open z kontekstem sięgającym miliona tokenów. Firma deklaruje przewagi w zadaniach agentowych, choć w testach kodowania model ustępuje części konkurentów.”
ai-sight.bsky.social, Bluesky · 18:47 UTC
Sources
13 articles from 13 outlets- The DecoderMuse Spark Open is Meta's answer to Qwen, GLM, and Kimi in the open-weight race
- The Verge AIMeta and Apple team up to make Muse AI models run better on Macs
- Hacker News front page (AI)Muse Spark Open
- QuartzMeta and Apple team up to run Muse Spark Open AI model on Macs
- StreetInsiderMeta releases Muse Spark Open, an 816B-parameter AI model
- Breakingthenews.netMeta, Apple team up to bring Muse AI model to Macs
- MacStoriesMeta Releases Muse Spark Open, an Over 800 Billion Parameter Model That Can Run Locally on Multiple M5 Ultra Mac Studios
- EngadgetMeta Built A Huge Open-Weight AI Model For Macs, But You'll Need 1TB Of RAM To Run It
- WSJMeta and Apple, Longtime Foes, Team Up to Put Open-Weight AI on Macs
- Axios"Mommy and daddy" play nice: Apple and Meta team up to bring open-source Muse to Macs
- 9to5MacMeta launches new Muse Spark Open model, optimized to run on Mac Studio
- Hugging Face BlogWelcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta
- ManilaShakerMeta Introduces Muse Wonder AI Models
Related Splits
All stories →
Anthropic posts $16.8 billion in Q3 revenue ahead of IPO
Anthropic reported third-quarter revenue of $16.8 billion, a 45% increase from the prior quarter, according to reports published between October 14 and 15, 2026. The company...
Nvidia and Apple each gave $25 million for Trump's White House ballroom
According to a Washington Post report published October 15, 2026, Nvidia and Apple each donated $25 million toward a $330 million fundraising effort connected to a...
Spy fears spread across Silicon Valley AI labs after OpenAI arrest
An arrest connected to espionage at OpenAI, reported on October 15 2026, has prompted broader concern across Silicon Valley AI laboratories about insider threats and foreign...