Meta teams up with old rival Apple to run its giant open AI model on two Mac Studios

Meta released the weights of Muse Spark Open, an 816-billion-parameter mixture-of-experts model it calls the most capable US open-weight model to date, on October 15, 2026, with Apple shipping a new version of its MLX framework on launch day so the compressed model runs across two Mac Studios with M5 Ultra chips and 512GB of memory each. Meta AI chief Alexandr Wang told CNBC the partnership follows "a long history between Meta and Apple," and the weights are free on Hugging Face, though a four-machine 256GB setup costs over $43,000 and Meta's own table shows Chinese models GLM 5.3 (45) and Kimi K3 (44) beating Muse Spark Open (39) on the Artificial Analysis Intelligence Index. The model has 42 billion active parameters per token, a 1-million-token context window, and comes in BF16 (about 1.6TB) and NVFP4 (about 530GB) versions, with AMD, Intel and Unsloth formats running on a 256GB machine via llama.cpp.

Two Mac Studios running Muse Spark Open on Mark Zuckerberg’s desk, in a photo he posted on X on October 15, 2026. Image: Mark Zuckerberg/X Meta has released the weights https://research.meta.ai/blog/muse-spark-open of Muse Spark Open, an 816-billion-parameter AI model it calls the most capable US open-weight model to date, and it worked with long-time rival Apple so the whole thing runs on two Mac Studios on a desk. Anyone can download it free from Hugging Face from today, October 15, though running it at home means about 1TB of memory and a hardware bill well into five figures. It is the open version of Muse Spark, the model behind Meta AI, and arrives three days after Meta launched its closed frontier model, Muse Wonder https://madrobot.blog/2026/10/12/meta-muse-wonder-frontier-ai-model-zuckerberg-price-availability/ . Mark Zuckerberg announced it on X with a nod to Apple’s new chief executive: Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple. — Mark Zuckerberg @finkd October 15, 2026 https://x.com/finkd/status/2110778025629347999 What Muse Spark Open is According to the model card https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B , it is a mixture-of-experts model with 816 billion parameters in total, of which only 42 billion are active for each token. It reads text, images, video and audio, writes text, and holds a context window of 1 million tokens. Meta built it for coding and agent work: long multi-step tasks, tool use and fixing real software bugs. The weights come in two versions from Meta: full-precision BF16 about 1.6TB and a compressed NVFP4 version about 530GB . AMD, Intel and Unsloth have made their own smaller formats, down to roughly 2-bit files that Meta’s developer guide https://dev.meta.ai/resources/blog/build-with-muse-spark-open says run on a 256GB machine with llama.cpp. The Apple team-up Meta and Apple have spent years trading blows over privacy and app store rules. This time Apple shipped a new version of its MLX framework on launch day and published its own benchmarks https://machinelearning.apple.com/research/mlx-distributed-inference . Two Mac Studios with M5 Ultra chips and 512GB of memory each, or four with 256GB, linked by Thunderbolt 5, hold the compressed model between them. Apple says a four-machine cluster generates more than 36 tokens a second, faster than most people read. Meta AI chief Alexandr Wang told CNBC https://www.cnbc.com/2026/10/15/apple-meta-partner-ai.html that the bad blood predates him: Most of the animosity predates my tenure at the company, but obviously there’s been a long history between Meta and Apple. Alexandr Wang, Meta AI chief, to CNBC The catch is cost. Apple has not yet priced the 512GB Mac Studio, which it says arrives in late October, and CNBC puts the 256GB M5 Ultra model at $10,799 or more in the US, so a four-machine setup costs over $43,000. Everyone else can rent it: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI host it, and the weights are also on Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. How good is it? “US” is doing real work in Meta’s headline claim. On Meta’s own table, Chinese models GLM 5.3 45 and Kimi K3 44 both beat Muse Spark Open 39 on the Artificial Analysis Intelligence Index, and Kimi K3 leads on several coding tests. Muse Spark Open does come top on some agent tests, such as MCP-Atlas 89.2 and JobBench 61.8 . The scores are Meta’s and haven’t been independently checked. The open version has extra safety training because, once downloaded, it runs without Meta’s guardrails. On the StrongREJECT jailbreak test, where lower is better, Meta reports a score of 0.5 against 62.9 to 86.4 for the rivals it lists. Wang was frank that this is not Meta’s best work: These aren’t the most capable models that we’ve developed or released. But we think that there will be a lot of companies and developers who sort of use this model and tinker with it at small scale and then seek to ramp up and scale up their usage. Alexandr Wang, Meta AI chief, to CNBC The licence Unlike the Apache-licensed Muse Glimmer https://madrobot.blog/2026/10/10/meta-muse-glimmer-how-to-run-locally-ollama-lm-studio-vs-qwen-gemma/ , Muse Spark Open uses Meta’s own Muse Model License. Commercial use, fine-tuning and distillation are allowed, but products must show “Built with Muse Spark Open”, and any company selling AI models as a service with more than $20 million in yearly revenue must sign a separate deal with Meta first. The licence covers users in Europe through Meta Platforms Ireland, and the model is not meant for under-18s. Why it matters This is the biggest model anyone has made a real effort to run on desktop hardware, and it gives Apple a selling point for its priciest Mac as it struggles to keep up in AI https://madrobot.blog/2026/09/25/apple-falling-behind-ai-race-siri-analysis/ . For most people the cheaper route is still a smaller local model on a Mac mini https://madrobot.blog/2026/09/23/how-to-run-local-ai-models-mac-mini/ , or renting Muse Spark Open from a cloud host.