Meta released an open AI model so big it needs 1TB of Mac memory

Meta released Muse Spark Open on October 15, an 816-billion-parameter open-weight model with 42 billion active parameters, a 1-million-token context window, and native multimodal support for text, images, video, audio, and PDFs, which requires roughly 1TB of unified RAM to run locally. Because Apple's top Mac Studio configuration with the M5 Ultra chip maxes out at 512GB of unified memory, running the model requires linking two maxed-out M5 Ultra Mac Studios over Thunderbolt, a setup that costs roughly $20,000 to $25,000. Meta calls Muse Spark Open the most capable open-weight model built in the United States to date, and the release lands amid rising memory prices, with server-grade 64GB RDIMM memory jumping from about $450 to over $900 between the fourth quarter of 2025 and the first quarter of 2026 and Goldman Sachs projecting a DRAM undersupply of roughly 4.9% this year.

Meta built an open-weight model tuned for Apple Silicon, then shipped specs that only two maxed-out Mac Studios, stacked together, can actually run. Meta released Muse Spark Open on October 15. It's an 816-billion-parameter model with 42 billion active parameters, a 1-million-token context window, and native multimodal support for text, images, video, audio, and PDFs, according to Engadget. Meta calls it the most capable open-weight model built in the United States to date. It is also, by Meta's own design, not something almost anyone can run at home. The catch is memory. Engadget reports the model needs roughly 1TB of unified RAM to run locally, which rules out every Mac Apple currently sells on its own. The top configuration of the new Mac Studio, built around Apple's M5 Ultra chip, tops out at 512GB of unified memory. Meta designed Muse Spark Open around exactly that ceiling, according to 9to5Mac: link two maxed-out M5 Ultra Mac Studios together over Thunderbolt and their combined memory pools form a single terabyte. That's enough to hold the whole model. That is not a cheap workaround. The M5 Ultra Mac Studio starts at $5,499 for a 96GB configuration. Apple's 512GB version, the one Muse Spark Open actually needs two of, doesn't even ship until late October, and industry pricing chatter already puts it well north of $10,000 per unit. Buy the pair and you are closer to $20,000 to $25,000 before you have run a single prompt, depending on how each machine is configured. Apple's unified memory architecture is the reason this is even possible at all. Unlike a discrete GPU with its own capped VRAM pool, Apple Silicon lets the CPU, GPU, and Neural Engine all draw from the same shared RAM. That's exactly why Meta picked the Mac as its target platform instead of a rack of GPUs. The timing makes the hardware math worse, not better. Memory prices are in the middle of what the industry has started calling "RAMageddon." Server-grade 64GB RDIMM memory jumped from about $450 to over $900 between the fourth quarter of 2025 and the first quarter of 2026. Goldman Sachs has projected a DRAM undersupply of roughly 4.9% this year, calling it the most severe memory shortage in over 15 years. The cause is AI infrastructure itself: data centers buying up HBM capacity are squeezing the same wafer supply that would otherwise make conventional DRAM. Meta is releasing a model that demands more consumer-grade memory than Apple has ever shipped in one machine, in the same year that memory got dramatically more expensive to produce. Mistral Releases Le Chonk, a 1 Trillion Parameter Open Weight Model https://startupfortune.com/mistral-releases-le-chonk-a-1-trillion-parameter-open-weight-model/ Simon Willison's test flagged a real gap: the API exposes only two reasoning levels, none and high, with nothing in between. Mistral also claims Large 4 leads AutomationBench at about a fifth of rival models' per-task cost. - mistral large 4 trillion parameter open weight model https://startupfortune.com/mistral-releases-le-chonk-a-1-trillion-parameter-open-weight-model/ - european AI lab releases trillion parameter model free https://startupfortune.com/mistral-releases-le-chonk-a-1-trillion-parameter-open-weight-model/ Muse Spark Open lands in the middle of a genuine arms race in open-weight models, and Meta isn't even the most extreme example. Mistral AI unveiled Mistral Large 4, nicknamed "Le Chonk," on October 6: a 1.05-trillion-parameter model with 49 billion active parameters and its own 1-million-token context window. Mistral has said it will release Le Chonk's weights on October 27 and is explicitly positioning the model as a challenger to China's leading open-weight systems. Nvidia, meanwhile, has gone a different direction. Its Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model with only 3 billion active parameters at a time. It's released under a license that permits commercial use and includes the training data and recipes alongside the weights. That split matters. Nemotron 3.5 Lightning is built to run on hardware a small team might actually own. Muse Spark Open and Le Chonk are built to prove a point about frontier capability. Running either one outside a cloud data center means buying into six figures of memory, whether that memory sits in GPUs or in a pair of Mac Studios. Open weights were supposed to mean you could download a model and own it. Increasingly, downloading it is the easy part. For Muse Spark Open specifically, "open" means Meta will hand you 816 billion parameters for free and then let the memory market decide whether you ever get to use them. Right now, for almost everyone outside an enterprise cluster, the honest answer is Meta's own cloud API, or nothing. Also read: Fed's Goolsbee warns the AI spending boom is now pushing broad inflation https://startupfortune.com/feds-goolsbee-warns-the-ai-spending-boom-is-now-pushing-broad-inflation/ • Nvidia Releases Nemotron 3.5 Super VL to Widen the Open-Weight AI Field https://startupfortune.com/nvidia-releases-nemotron-35-super-vl-to-widen-the-open-weight-ai-field/ • CoreWeave default odds hit 60 percent as credit swaps top 1,000 basis points https://startupfortune.com/coreweave-default-odds-hit-60-percent-as-credit-swaps-top-1000-basis-points/ This article is posted in AI News https://startupfortune.com/category/ai/ , check it out for more related stories. Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.