On 12 October Meta priced Muse Wonder into the frontier API bake-off. Three days later the same lab published downloadable weights for Muse Spark Open and showed the model running across a pair of Apple Mac Studios, wired together like one machine. That desk demo is the signal: Meta wants a large open-weight model to feel local, and Apple’s M5 Ultra silicon is the showroom.
Meta and Apple split an 816B MoE across two M5 Ultra Mac Studios
Muse Spark Open is a mixture-of-experts model with 816 billion total parameters and 42 billion active per token. Context reaches 1 million tokens. Inputs include text, image, video and audio; output is text. Meta’s research post dates the weight release to 15 October 2026 and frames the model for coding and long agent runs.
The hardware claim is unusually specific. Meta’s platform docs say the NVFP4 checkpoint runs on two Mac Studio computers with M5 Ultra and 512 GB of unified memory each, or four with 256 GB each, through Apple’s MLX stack. MLX shards the model over RDMA on Thunderbolt 5, which needs macOS 26.2 or later and a direct link between every pair of Macs. Meta’s launch copy says the two Studios behave like one terabyte machine, with enough bandwidth to keep feeding the active slice of the MoE.
Weights are listed on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted partners named at launch include Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, with runtimes spanning SGLang, vLLM, TensorRT-LLM, MLX and others across NVIDIA, AMD, Intel and Apple silicon.
Meta showed Muse Code, powered by Spark Open, shipping real changes to Ice Cubes, an open-source Mastodon client, and driving DaVinci Resolve to review footage and crop vertical clips. Those are company demos; Meta left token-per-second figures for the two-Studio setup unpublished.
Mark Zuckerberg thanked Apple CEO John Ternus on X. Ternus congratulated the team for running a model of that scale “right from your desk.” Alexandr Wang told CNBC and Axios that Meta had worked with Apple for months, with former Apple executive Daniel Gross helping open the channel.
Desktop agent builders and US open-weight buyers feel it first
The first audience is practical: teams that want a US-developed open-weight model they can download and fine-tune, Mac shops already living in MLX that can afford a multi-Studio rack, and counsel checking whether their product counts as model-as-a-service under the Muse Model License.
Most buyers will start on a host. The 512 GB Mac Studio Meta leads with had not reached Apple’s store at announcement; MacRumors and CNBC put it later in October, with price still unlisted. A 256 GB M5 Ultra was already listed from about $10,799, so four machines for a terabyte start in the low five figures before tax and links.
Meta adds a downloadable Muse weight file to a priced API week
Wonder, released on 12 October, is Meta’s frontier API and Muse Code subscription model, priced into the Opus-class bake-off. Spark Open lands beside it as a downloadable weight file, now demoed on a Mac desk. Wang told CNBC that Spark Open sits below Meta’s most capable releases, and that the bet is developers who tinker locally and then scale usage.
Meta calls Spark Open the most capable US open-weight model to date. The geographic word matters. Chinese open-weight systems still set much of the download and coding pace, and Hugging Face and OpenRouter traffic figures cited in secondary coverage have made that hard to ignore. A US weight file that runs on a desk gives procurement teams a domestic option with local control.
Meta’s own launch table, as reported by VentureBeat, puts Spark Open ahead on JobBench and MCP-Atlas tool use, and behind Kimi K3 and GLM 5.3 on Terminal-Bench, DeepSWE and GDPVal-AA. The Artificial Analysis Intelligence Index score on Meta’s card is 39, below several Chinese open models’ independently published mid-40s scores. Until Artificial Analysis posts its own Spark Open row, treat that 39 as Meta’s figure.
Together AI listed serverless Spark Open at $1.25 per million input tokens, $0.15 cached and $4.25 output, matching Meta’s hosted Spark family rates and well below Wonder’s $4/$20. Host prices will vary; Groq’s docs already note org approval and a shorter context cap than the full 1M window.
Coding rows, unshipped memory, and a MaaS license still decide the buy
“Most capable US open-weight” is a category Meta can defend on its sheet. On your repository, Kimi or GLM may still win the coding harness. DeepSWE rows on Meta’s table use each model’s preferred agent stack, so the surrounding software is part of the score, and independent reproduction of the desktop demos and Index line was still thin in the first coverage wave.
The Muse Model License is open-weight with conditions, including “Built with Muse Spark Open” attribution. The commercial gate that drew attention applies when a licensee and its affiliates offer third parties API-style access to AI models and their aggregate gross revenue exceeds $20 million in any consecutive twelve months. Then Meta wants notice and a separate written agreement before that model-as-a-service business can offer Spark Open or derivatives. Secondary reads put internal use, research and in-product features outside that rule.
Hardware timing softens the romance. Two 512 GB Studios are the poster configuration, and that SKU was still forthcoming at launch. Meta’s model card also lists heavy accelerator memory for GPU serving, on the order of 700 GB for NVFP4 and far more for BF16, so holding the full MoE still takes a large rack even when only 42B parameters are active per token.
Meta says Spark Open took extra open-weight post-training, met its Superintelligence Scaling Framework release thresholds, and will publish a fuller evaluation report soon. Until that report has a public URL, the safety claim stays incomplete.
Try the hosted path before you buy a pair of Studios
API and eval teams can put Spark Open on one multi-hour coding or tool-use job this week through a host they already bill, keep Kimi or GLM in the same harness, and let the traces settle the ranking. Mac-native shops should wait for the 512 GB Studio price and a measured tokens-per-second note before treating two Studios as the default rack; four 256 GB units are already an option if privacy and offline work justify the capital. Product counsel should map whether they sell model access as a service and whether affiliate revenue clears $20 million.
Meta and Apple have made a large US open-weight model feel like something you could run from a desk. The open questions are whether those weights catch Chinese coding leaders on a shared harness, and whether anyone signs a purchase order for a terabyte of Mac memory before the hosted bill looks cheaper.