Meta released Muse Spark Open on October 15, 2026, and the headline number is 816 billion parameters. I run Ollama with Nous Hermes 4 for client-sensitive drafts, so my first question was whether this model could ever sit next to it. I have not run it. What I did was read Meta’s announcement and the Hugging Face launch post line by line and pull out the five numbers that decide whether a one-person shop should care.
In this article
- Why 816B parameters only means 42B per token
- What Muse Spark Open costs in memory on a Mac
- Why the 1M-token window is a server feature
- How always-on reasoning changes the wait
- What the license allows, and my plan
Only 42B of the 816B parameters work on each token
The size sounds absurd, but the architecture is a mixture-of-experts design. According to the Hugging Face post, each layer holds 256 experts and routes every token to 8 of them, so about 42B parameters are active per token. The Hub lists the total at 822B while the repo name says 816B, a small discrepancy worth knowing before you search for files.
That helps speed, not storage, and it is the first thing to understand about Muse Spark Open. All the experts still have to sit in memory, because any token can call any of them. Experts hold roughly 97% of the parameters, so shrinking the active slice does little for your disk or RAM.
Here is what the Hugging Face post says about the inputs and the vision side:
- Inputs are text, images, audio, and video, and the output is text only.
- Video is sampled at 2 frames per second by default, capped at 96 frames.
- Audio is mono at 24 kHz, with one audio token every 80 milliseconds.
For a consultant, that means Muse Spark Open could read a screen recording or a call clip, but it will not draw you a slide. Meta describes it as built for coding and agentic work, and calls it “the most capable US open-weight model to date.”
Muse Spark Open needs about 304GB even at 2-bit on a Mac
The most useful numbers in the launch post come from a Mac M3 Ultra with 512GB of memory. With llama.cpp at Q2_K, the model uses about 304GB and generates roughly 18 tokens per second. With an MLX 3-bit quant, it uses about 344GB and generates roughly 21 tokens per second.
Those are the best-case consumer numbers, and they come with caveats:
- The MLX run used a 1,024-token prompt and 128 generated tokens, a short test.
- Prefill ran at about 225 tokens per second, so long documents wait before the first word appears.
- The post says the 3-bit MLX quant needs more than 256GB of RAM.
Meta’s own page mentions two Apple M5 Ultra Mac Studios running the model through MLX. My earlier Mac Studio M5 Ultra memory breakdown covers why 512GB matters, and this model is the first case where it is the floor rather than a luxury.
Here is the arithmetic I did myself from those figures. A 512GB machine minus 304GB of weights leaves about 208GB, and a full 1M-token cache needs about 266GB. Even the best consumer setup in the post cannot hold the full window, so the practical local context is much shorter than the headline number.
The 1M-token context is a server feature, not a laptop feature
The model supports up to 1,048,576 tokens of context. The Hugging Face post puts the KV cache at about 250KB per token in BF16, which works out to about 266GB for a full window. That is on top of the weights.
The official BF16 checkpoint is about 1.5TB. The post’s reference setup for it is four nodes of eight H200 GPUs. The 4-bit NVFP4 checkpoint fits a single B200 node, but it targets NVIDIA Blackwell hardware.
“The 1M-token window is real, but filling it needs 266GB of cache before you load a single weight.”
So a solo operator has two realistic paths: a heavily quantized local run with a short context, or a hosted API. Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI as hosted options, and gives no pricing on its page.
Reasoning you cannot turn off changes the wait
This is the detail I would have missed from the headline alone. The Hugging Face post lists four reasoning effort levels (low, medium, high, xhigh), says the default is xhigh, and states that “turning off reasoning is not possible with Muse Spark Open.”
For quick client emails, that matters. At roughly 18 to 21 tokens per second locally, any hidden reasoning tokens add directly to the wait before the answer starts. My own back-of-envelope estimate: a 1,000-word draft is around 1,300 tokens, so about a minute at those speeds, before any reasoning overhead. I would call that a planning guess, not a measurement. If your day is mostly short replies and quick rewrites, the always-on reasoning makes this a poor fit for that rhythm, while a long agentic coding run is closer to what Meta designed it for.
For short drafting jobs, I would rather keep my current Ollama setup and treat Muse Spark Open as a second tool for heavier work. My Muse Glimmer read from August covered Meta’s much smaller on-device model, which suits that job better.
The license is free for many developers, not all
Meta releases the weights under the Muse Model License and describes it as having “limited commercial terms” that “allow many developers to use the model for free.” The word “limited” is the part to read before building anything billable on it. I have not seen the full terms, so I will not guess at the thresholds.
Two other gaps are worth noting:
- Meta’s page has a Key Benchmarks section but lists no scores, pointing to an evaluation report that is “available soon.”
- The Hugging Face post does not mention Ollama, so I cannot tell you it will run there yet.
My plan: wait for the benchmark report before downloading Muse Spark Open
For me, Muse Spark Open is a watch item rather than a download. The memory floor is more than 300GB, the reasoning cannot be switched off, and the benchmark report that would justify the hardware has not shipped. I would revisit it once independent scores and the full license text are out.
Before any download, I would check four things in order:
- The independent benchmark scores, once Meta’s report is out.
- The full Muse Model License text, against my own client contracts.
- Whether Ollama or another runner I already use supports the model.
- Whether a hosted API’s data terms would let me test with anything beyond public material.
If I did test Muse Spark Open, I would start with a hosted API on public, non-client material only, which fits my local AI privacy rules for client data. Local use would come later, if at all, and only on hardware I already own.
Sources
AI-assisted research and drafting. Reviewed and published by ToolMint.