Meta has released the weights for Muse Spark Open, a very large AI model for coding and agent work that developers can download and run on their own hardware. Apple, working with Meta, says the latest version of its MLX software can run the model across a small cluster of Mac Studio desktops, generating more than 36 tokens per second on four machines.
Meta calls Muse Spark Open “the most capable US open-weight model to date,” according to its launch blog post as quoted by 9to5Mac. That is Meta’s own claim. The model’s weights are public, but it is not fully open source: CNBC reported that it is released under a licence that allows only limited commercial use.
What the model is
Meta’s developer documentation describes Muse Spark Open as “Meta’s 816B-parameter mixture-of-experts model with 42B active parameters, for agentic and coding work, with text, image, audio, and video input.” A mixture-of-experts model uses only part of its network for each word it produces, which cuts the computing needed. The weights ship in two formats, full-precision BF16 and a compressed 4-bit format called NVFP4.
Meta also says the model has a 1M-token context window, meaning it can take in very long documents or codebases at once. That figure, too, comes from Meta.
Open weights mean developers can run the model without sending data to Meta. Meta’s documentation notes that with open-weight models “you own the weights, the runtime, and the data.”
How it runs on Macs
Apple’s machine learning team set out the details in a research post on 15 October 2026, the same day it released a new version of MLX, its open-source framework for machine learning on Apple silicon.
The model is too big for any single Mac. Apple says it would need “more than 1.6TB of RAM” in BF16 for the parameters alone, and about 493GB in NVFP4. Once working memory is added, “the model cannot fit on a single machine,” even a Mac Studio with 512GB. Instead, Apple links machines over Thunderbolt using a driver for RDMA, a technique that lets computers read each other’s memory directly.
Apple’s own benchmarks, run on the NVFP4 version with 8,192 tokens of context, show:
- Two Mac Studios, pipeline mode: 15.74 tokens per second.
- Two Mac Studios, tensor mode: 24.99 tokens per second.
- Four Mac Studios, tensor mode: 36.73 tokens per second, “well above average human reading speed,” Apple says.
Apple also says the M5 Ultra is up to 4.6 times faster than the M3 Ultra at time to first token on other models it tested, and that each M5 Ultra Mac Studio peaks at about 300 watts.
Meta’s blog, as quoted by 9to5Mac, says two Mac Studios “behave like one supercomputer with a terabyte of unified memory.” Meta said it tested agent workflows entirely on-device with its Muse Code tool, including making changes to an open-source iOS app and editing video in DaVinci Resolve.
A rare Apple-Meta partnership
CNBC reported that this is the first public partnership between the two companies since 2017, when Apple removed system-level Facebook integration from the iPhone. Meta AI chief Alexandr Wang told CNBC that many powerful open-weight models are too big for most personal computers, but Mac Studios have enough memory to make running Muse Spark Open practical.
According to CNBC, Muse Spark Open is not as advanced as Muse Wonder, Meta’s proprietary model released earlier this week.
It won’t be cheap
The hardware is costly. CNBC reported that a Mac Studio with the required M5 Ultra chip and 256GB of memory costs at least $10,799, a 512GB version is due later this month, and users need several machines.
What’s still unknown
The full licence terms, including what “limited commercial use” covers, have not been independently reviewed by TSN. Independent tests of Meta’s capability claims and its 1M-token context are still to come.
What to watch
Watch for independent benchmarks, the 512GB Mac Studio’s launch and how developers use the model for local coding agents.
What this means
- Frontier-scale AI at a desk. A model of 816B parameters can run on a few Mac Studios.
- Open weights, not open source. CNBC reports the licence allows only limited commercial use.
- Claims to test. “Most capable” and the 1M-token context are Meta’s own claims.
Sources
- Apple Machine Learning Research, “Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio”, 15 October 2026
- Meta Model API documentation, “Open-weight models”
- CNBC, Meta and Apple partnership on Muse Spark Open, 15 October 2026
- 9to5Mac, “Meta launches new Muse Spark Open model, optimized to run on Mac Studio”, 15 October 2026
Related stories
- Meta Launches Muse Wonder and Muse Wonder Full, Its “Most Capable Models to Date”, Built for AI Agents
- AMD Launches Its Next Ryzen AI Halo Developer Platform, With Up to 192GB of Memory for Running Large AI Models Locally
- NVIDIA Readies Its RTX Spark Laptops for Local AI Agents With a New Open Model, Nemotron 3.5 Bolt
- Apple Finally Approves External GPUs for Mac — But Only If You’re Training AI, Not Gaming
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
