Meta released Muse Spark Open this week: 816 billion total parameters, 42 billion active, multimodal, 1M-token context, open weights. MarkTechPost led with the benchmark line, that it tops open models on MCP-Atlas. Hugging Face ran the welcome post with the usual integration details. Both are accurate and both buried the number I keep coming back to, which is that Meta says it runs on two M5 Ultra Mac Studios.
That is a strange flex for an 816B model, and it’s the most informative thing in the release.
Sparsity is the whole product
A 42B active slice out of 816B total is roughly a 5% activation ratio. The weights still have to live somewhere, so you need the memory to hold them, but the compute per token looks like a mid-size dense model. Unified-memory Macs are bad at raw FLOPS and good at holding enormous tensors with decent bandwidth. Muse Spark Open is shaped to exploit exactly that mismatch.
So the deployment story changes shape. Not cheaper, two maxed Mac Studios is not a hobby purchase, but it’s capex you buy once and keep in a closet, rather than an H100 cluster you rent by the hour from someone who logs your prompts. For anyone running agents over private data, that distinction is the entire reason to care about open weights in the first place.
MCP-Atlas is the benchmark that matters here
Leading open models on MCP-Atlas means tool-calling and protocol-following, not trivia recall. Combine that with a 1M-token context and the intended use is obvious: long-running agents that hold a large codebase or document corpus in context while calling out to real tools. Meta is not pitching a chatbot.
I’d treat the leaderboard position as a floor rather than a verdict. MCP-Atlas scores have a short shelf life, and “leads open models” is a narrower claim than it sounds, since the field it leads moves every few weeks. What won’t move is the architecture decision, and that’s the part other labs will copy.
What I want to know
Neither source tells us the per-token latency on that two-Mac setup, and that’s the number that decides whether this is a usable agent host or a demo. Sparse MoE routing across a 1M-token context is memory-traffic hell at long sequence lengths; prefill on a 500k-token prompt could be minutes. Nobody has published those curves yet.
The quantization question is also open. Two M5 Ultras at maximum configuration is a lot of unified memory, but 816B parameters at anything above 4-bit starts getting tight, and MoE models have historically been fussy about aggressive quantization because the routing gets noisy.
Still, the thing is downloadable, the license is open-weight, and the activation ratio means a serious local agent stack no longer requires renting datacenter silicon. Two Mac Studios and the patience to wait for the download.