Meta opens 816B Muse Spark weights — and two Macs can run them
Meta put a frontier-scale model in the open and Volvo handed Nvidia its whole driving stack. Google, meanwhile, is trying to make up with the publishers it spent two years starving.
Meta released the weights for Muse Spark Open, an 816-billion-parameter multimodal model it calls the most capable US open-weight release to date — and it runs locally on a pair of M5 Ultra Mac Studios. The mixture-of-experts model activates 42 billion parameters per token and carries a 1-million-token context window, taking text, images, audio, and video as input while generating text, which makes it the first American open-weight model aimed squarely at long-horizon agentic coding on hardware you own. Meta says it worked with Apple on the MLX optimizations — expert routing, Thunderbolt clustering, native multimodal support in mlx-lm — that let two to four Mac Studios hold the full model, roughly 1.5 TB in BF16. Zuckerberg and Apple CEO John Ternus traded congratulations on X, which would have been unthinkable when the two companies were feuding over ad tracking.
The bill is real, and so is the fine print. A Mac Studio with the required M5 Ultra chip and 256 GB of memory starts at $10,799, and you need a second one; Engadget puts the practical bar at about a terabyte of RAM. The weights ship under the Muse Model License with limited commercial terms, a license Nathan Lambert is blunt about — "the Meta Muse Open License is trash," he wrote, predicting providers will comply anyway. Meta's own framing is honest about the trade: Alexandr Wang said the open release isn't among "the most capable models that we've developed or released," but that it should pull developers in at small scale and let them ramp. Our take: the interesting number isn't 816 billion — it's that Meta shipped the closed frontier model first and the open one a week later, which is the reverse of how labs usually treat their best work. Open weights are becoming the second-tier product line, not the morale project. We covered the flip side of this hardware story in September — M5 Ultra reviews: local AI's ceiling is memory, not compute.
Volvo will put Nvidia's Hyperion compute and sensor architecture, Alpamayo driving software, and Halos safety layer into its future cars for a Level 2++ hands-free system that handles traffic lights, stop signs, and unprotected left turns. The automaker says point-to-point driving arrives "before the end of the decade," across the whole lineup rather than EVs alone, and without lidar for now. Nvidia runs Alpamayo's AI models alongside a classical perception stack at the same time, reasoning that redundancy is the safety case; Nvidia's automotive chief Ali Kani says the open-source Alpamayo software has passed 650,000 developer downloads. Read it against Volvo's history: its unsupervised Ride Pilot never got US regulatory approval, and its lidar supplier Luminar went bankrupt, so this is a company buying a second chance rather than extending a working one.
Google is moving inline links higher into its AI answers, part of a slow retreat from the link-free results that publishers have spent two years protesting. Google says inline links will appear "higher up in responses more frequently" in AI Mode and AI Overviews, is testing contextual links at the top of a response for some queries, is rolling its link carousels out to more languages, and is testing clickable links into publisher search profiles. It also says more than 400 publishers are now in its News AI licensing pilot, with Axios, CNN, and the Financial Times newly onboard. None of this reverses the traffic problem — a citation at the top of an answer still isn't a click — but it is the first time Google has conceded the format itself was the issue rather than the volume of links.
What to watch: whether anyone outside Meta reproduces its agentic benchmark numbers, and whether Volvo's "before the end of the decade" firmness survives a single product cycle.
Meta shipped the weaker model open and kept the strong one closed. Is open-weight enough, or does it just mean everyone gets yesterday's frontier? Tell us in the comments.
