Meta just put the weights of its biggest model yet online. The interesting part isn't the size. It's that a demo ran it on two Mac Studios.
Meta opens up Muse Spark after five months of keeping it closed
Meta released the weights for Muse Spark Open on October 15, turning a model that had been proprietary since its April debut into something anyone can download. The file lives on Hugging Face under a new Muse Model License. Meta calls it the most capable open-weight model to come out of a US lab so far.
That's a company talking about its own work, so treat the phrasing as a claim rather than a fact. What's not in dispute is the shape of the thing. Muse Spark Open is a mixture-of-experts model, which means it's built from many specialized sub-networks and only wakes up the parts a given prompt needs. It holds 816 billion parameters in total, but just 42 billion of them fire for any single token. The context window, the amount of text and media the model can consider at once, runs to one million tokens.
Input isn't limited to typing. Prompts can mix text with images, audio, or video. Meta built this one for coding and long-running agent work, the kind of jobs where a model has to keep its bearings across dozens of steps.
What Muse Spark Open can actually handle
Meta says training leaned on what it calls real developer workflows, including multi-file refactors and long debugging sessions. That's the pitch, and it points at a specific audience: developers who want a capable model they can run themselves instead of renting through an API.
The model started from Muse Spark 1.2, which shipped in August. Meta then added a round of safety post-training so the release could stand on its own without the guardrails that wrap Meta's own services. The company says it cleared the thresholds in its Meta Superintelligence Scaling Framework, with a full evaluation report still to come.
Two weight formats are on offer. One is the standard BF16, the higher-precision option. The other is NVFP4, a compressed format tuned to run on NVIDIA and AMD hardware. SGLang and vLLM, two popular tools for serving models, supported it from day one.
The Mac Studio demo is the part worth pausing on
Running an 816-billion-parameter model usually means a rack of data-center GPUs. Meta ran a demo on two Apple Mac Studio desktops with M5 Ultra chips, linked into a single system through Apple's MLX framework. Between them, the two machines had about a terabyte of memory.
On that setup, Muse Code, Meta's coding agent, worked through bug fixes and new features for Ice Cubes, an open-source Mastodon client. A second test pushed video from DaVinci Resolve through the model frame by frame to pick and crop the best shots for vertical clips.
The takeaway for a normal developer isn't that you'll be running this on your desk. It's that the ceiling for local hardware keeps rising. A studio that already owns high-end Macs might soon be able to prototype against a model in this class without booking cloud time. That's a shift worth tracking.
Meta's benchmark numbers versus the wider field
Meta's own benchmark table puts Muse Spark Open at 61.8 on JobBench, an agent test, which the company says is more than double the 31.1 scored by Inkling from Thinking Machines Lab. NVIDIA's Nemotron 3 Ultra trails at 27.2 in the same table. Those are Meta's figures, measured and reported by Meta.
Coding is a different story. On Terminal-Bench 2.1, which tests command-line work, Chinese open models sit ahead. Kimi K3 leads at 85, and both GLM 5.3 and Qwen3.8 finished above Muse Spark Open's 76.4.
Tool use flips the order back. On MCP-Atlas, Muse Spark Open scored 89.2, ahead of every rival Meta listed. That benchmark hands a model a plain-language task without naming the tools it needs, then asks it to pick and sequence the right ones from 220 tools across 36 live MCP servers. Getting that right matters more than raw coding speed for anyone building agents that touch real systems.
What the Muse Model License lets you do
The weights come with strings. Meta describes them as limited commercial terms. Developers who adapt the model or ship it inside their own software pay nothing. So what trips the fee? Companies that sell it as a hosted service and have passed $20 million in gross revenue over a 12-month period. Those need a separate agreement.
Products built on the model have to carry a "Built with Muse Spark Open" label. Military and critical infrastructure uses are off the table entirely. If your plan is to resell access, read the license before you build.
Where to get Muse Spark Open
Seven inference providers are offering hosted access at launch, including Groq and Fireworks AI. You can also pull the weights straight from Hugging Face or from the model catalogs of AWS, Google, and Microsoft.
The bigger picture: Meta's most capable systems stay closed. The Muse Wonder models it introduced earlier this week are reachable only through the Meta Model API and Muse Code. Muse Spark Open is Meta's bid to lead the open-weight race it once sat out, with the training run behind the last open release, the 30-billion-parameter Muse Glimmer, now dwarfed by comparison.
