Muse Spark Open: Meta's 816B Model You Can Run Yourself
 |
Muse Spark Open: Meta's 816B Model You Can Run Yourself
Meta put the weights of an 816-billion-parameter coding model on the internet for free — and the headline demo ran it on two Apple desktops instead of a data center.
On October 15, Meta released the open weights of Muse Spark Open, an 816-billion-parameter model built for coding and agentic work — long, multi-step tasks where an AI uses tools, writes code, and fixes its own mistakes. Anyone can download it from Hugging Face right now, free, and Meta is calling it the most capable open-weight model to come out of a U.S. lab.
But the strangest part wasn't the release. It was the demo: Apple helped Meta optimize the model to run on two M5 Ultra Mac Studios linked together — consumer desktops, not a server rack. Two companies that spent years publicly fighting over privacy and the App Store now have a joint AI showcase.
So what did Meta actually release, how can an 816-billion-parameter model possibly run on a desk, and should you care?
How 816B fits on a desk: only 42B wakes up
Muse Spark Open is a mixture-of-experts model — think of it as a model built from 256 smaller specialist sub-networks ("experts"), where a router wakes up only the 8 experts each token needs. Every expert still has to live in memory, but each token is computed with roughly 42 billion active parameters instead of 816 billion. That's the trick that makes it fast.
The full-precision BF16 weights are about 1.6 terabytes; a compressed 4-bit NVFP4 build shrinks to roughly 493 GB. But Apple's researchers say that still doesn't fit on one Mac Studio once you add the working memory the model needs while answering — so you need two machines with 512 GB of unified memory each, or four with 256 GB, linked over Thunderbolt 5 with RDMA, which lets one computer read another's memory almost directly.
Context goes up to 1 million tokens — roughly a small library's worth of code in one prompt. The model takes text, images, audio, and video as input and returns text. It follows Muse Glimmer, the 30-billion-parameter dense model Meta released in August, and delivers on the open-weights promise Meta made for the Muse Spark family back then.
A reasoning_effort setting (low to xhigh, default xhigh) dials how hard the model thinks — but you can't switch reasoning off, so every answer burns extra tokens.
Apple's benchmarks: how fast is "two desktops"?
Apple actually measured this. On the 4-bit build at 8,192-token context:
- Two Macs, tensor parallelism: 24.99 tokens per second, 8.2 seconds to the first token
- Two Macs, pipeline parallelism: 15.74 tokens per second, 8.4 seconds to the first token
- Four Macs, tensor parallelism: 36.73 tokens per second, 4.8 seconds to the first token
Tensor parallelism — splitting every layer across the machines — is what scales as you add Macs; pipeline parallelism mostly helps on very long prompts.
The demos were real agentic work, not chat: Muse Code powered by the model shipped actual iOS changes to Ice Cubes, an open-source Mastodon client, and drove DaVinci Resolve end to end, picking the strongest moments out of footage.
The honest benchmark read
Meta published its own comparison table, so treat every number below as vendor-selected — but it's still informative, especially where Muse Spark Open loses.
It shines at agentic tool use: 89.2 on MCP-Atlas (ahead of Qwen3.8's 84.5 and GLM 5.3's 84.2) and 61.8 on the JobBench agent test — more than double Inkling's 31.1. That's why Meta is pitching this as a coding-agent model rather than a chat model.
Raw coding is the weak spot. On Terminal-Bench 2.1 it scored 76.4 — behind Kimi K3 (85.0), GLM 5.3 (83.9), and Qwen3.8 (82.0). Its Artificial Analysis Intelligence Index of 39 trails GLM 5.3 (45), Kimi K3 (44), and Qwen3.8 (40).
The word "U.S." is doing real work in Meta's headline claim — its own table admits the strongest open-weight coding models right now come from Chinese labs.
Wang himself undercut the headline. "These aren't the most capable models that we've developed or released," Meta's AI chief Alexandr Wang told CNBC. "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." It's a gateway drug to Meta's paid models, and he's not pretending otherwise.
Not open source. Open weights are not the same thing. Muse Spark Open ships under Meta's own Muse Model License: use, fine-tuning, distillation, and redistribution are allowed, but products must display "Built with Muse Spark Open," and any company selling model access as a service with over $20M in yearly revenue needs a separate agreement with Meta. Free to download isn't free of strings.
What it would actually cost you to run
Here's the part the demo skips. The 512 GB Mac Studio in Meta's demo isn't even shipping yet — it's due later this month — and the 256 GB version already costs over $10,000. Cost estimates put a pair of the high-memory machines at $30,000 to $40,000; Engadget estimated around $38,000 for the four-machine 256 GB setup.
That makes the hosted route the sane default: the weights are live on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, with OpenRouter listing $1.25 per million input tokens and $4.25 per million output tokens. Below a very large monthly volume, the API beats the hardware on cost — a DEV Community breakdown of the real numbers reaches the same conclusion.
Day-one inference engines include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, and llama.cpp. On a budget, Unsloth's 2-bit GGUF builds reportedly squeeze it into about 225–275 GB — still too big for a laptop, but within reach of a fat workstation.
What trips people up
Confusing "42B active" with "small." Active parameters are about speed, not size: all 816 billion still have to sit in memory. The experts don't disappear just because most of them are sleeping.
Reading Meta's benchmark table as a neutral ranking. The interesting signal isn't the wins — it's that Meta published losses too, then had its AI chief publicly undercut the headline. Read the table as a positioning document, not a leaderboard.
Ignoring the reasoning tax. With reasoning effort defaulting to xhigh and no off switch, the model burns extra output tokens thinking through every answer. At $4.25 per million output tokens, that thinking costs real money on the API — and real seconds on a local setup.
For most developers, the move is simple: try it on a hosted API first — it's a genuinely strong open-weight agent model for tool-using workflows. Think about hardware only if your usage is huge, your data can't leave the building, or you want to own the biggest model your desk can hold. The ceiling for open weights just moved — and the "not our best model" honesty is part of why this one is worth watching.
Sources: Reuters; Meta's model card on Hugging Face; Brocker.org; Teck Hustlers (Apple benchmarks, hardware detail); madrobot.blog (benchmark table, Wang CNBC quote); DEV Community (hardware reality); DEV Community (self-host cost math); Implicator (Mac-vs-cloud cost).
Where can I download the Muse Spark Open weights?
From Hugging Face (BF16 and NVFP4 checkpoints), plus Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. You'll need a Hugging Face account and to accept the Muse Model License first.
Did Meta make the open version safer than the closed one?
That's the claim. Because the downloaded model runs without Meta's guardrails, Meta says it gave this release extra safety post-training — and on the StrongREJECT jailbreak test it reports 0.5 against 62.9–86.4 for the rivals it lists. One caveat: that's Meta measuring its own model.
Can I buy the 512 GB Mac Studio yet?
Not yet. Apple has promised it for late October 2026; at this writing the 256 GB machine — already over $10,000 — is the biggest Mac Studio you can actually buy. The four-machine 256 GB setup is the path that exists today.
