Can You Run Muse Glimmer on the Mac You Already Own? The Real Memory Math

Quick answers, for the impatient:

• Can I run Muse Glimmer on a 16GB Mac? No. The default download is 18GB and a 16GB Mac gives Metal about 12.1GB to work with. It will not load.
• What are Muse Glimmer’s real memory requirements? 24GB of unified memory at the absolute floor; 32GB to be comfortable. Independent reviewers land on the same range.
• Will it run on a Mac mini? A 32GB Mac mini, yes — one owner’s verdict: it works, but “everything runs slowly.”
• Is there a smaller version? No. Every Ollama tag is the same 30B model, just quantized differently.

Now the long version, with receipts.

What Muse Glimmer actually is

Meta released muse-glimmer on Ollama in August 2026, and the spec sheet is genuinely attractive: a 30B dense model with a 1.8B vision encoder for perception, Apache 2.0 licensed, 128K context, text and image input, and first-class tool-calling / function-calling support. It’s distilled from “Muse Spark,” and Meta’s benchmark story is all about agents: 75.5% on MCP-Atlas, 51.2% on SWE-Bench Pro, 74.6% on DeepSearch QA.

The pitch writes itself: a capable, permissively licensed agent model running on your own machine, no API bill, no rate limits.

The part that isn’t in the launch post: almost nobody’s existing Mac can load it.

The download sizes: every tag is 30B

Here is what Ollama will actually hand you, by quantization:

Note what’s missing: a smaller model. There is no 8B or 3B parameter variant of Glimmer. Every single tag is the same 30B dense network. Your only lever is quantization, and the smallest official quant is 18GB.

If your mental model is “18GB model, 16GB Mac, that’s close, maybe it squeezes” — it doesn’t, and the reason is worse than the headline numbers suggest.

The 16GB Mac problem: Metal gives you 12.1GB, not 16

This is the number that most coverage skips, and it’s the one we measured ourselves.

On a 16GB Apple M4 Mac running macOS 26.4, we queried Metal’s actual GPU memory budget. The answer — recommendedMaxWorkingSetSize — came back as 12,713,115,648 bytes, roughly 12.1GB. That’s what macOS is willing to let the GPU use out of your 16GB of unified memory. The rest is reserved for, well, running your Mac.

So the real comparison on a 16GB machine isn’t 18GB vs 16GB. It’s 18GB vs 12.1GB. The default quant misses by nearly six gigabytes, before you’ve allocated a single token of context.

And this lines up with the independent coverage: both Notebookcheck and kingy.ai put the practical minimum at 24–32GB of unified memory or VRAM, and the press verdict is blunt — “a typical 8GB or 16GB laptop remains out of reach.”

What happens when you try anyway

Two things, and neither is the fun kind of hacking.

Ollama refuses to load it. Ollama checks model size against available memory and declines to load models that exceed it, rather than swapping to disk. On our 16GB M4, the 18GB default quant simply cannot load. This is honestly the polite failure mode — you get a refusal instead of a frozen laptop.

Forcing CPU inference is worse than nothing. If you route around the GPU and run a 30B dense model on CPU with insufficient RAM, the machine page-thrashes — constantly shuttling model weights between RAM and disk — and generation speed lands at well under 1 token per second. That is not “slow but usable.” At under a token per second, a single agent tool call is a coffee break, and an agent loop of a dozen calls is an afternoon.

Speed reality on Macs that can run it

Suppose you own, or buy, a Mac with enough memory. What do you actually get?

• M4 Pro-class chips: roughly 10–14 tokens/second text decode, per independent testing by kingy.ai. Usable for chat; leisurely for agent work.
• M4 Max with DFlash, Meta’s speculative-decoding drafter: 23.7 tok/s baseline rising to 37.8 tok/s — about a 1.5x speedup.
• M5 Max with DFlash: 26.6 tok/s rising to 50.2 tok/s — about 1.8x.

So the good news is real: on a Max-class chip with the drafter, Glimmer moves at genuinely workable speeds. The catch is that “Max-class chip with lots of unified memory” is precisely the Mac most people don’t already own.

And the floor of “can run it” is humbler than the benchmark charts. A Hacker News user running Glimmer on a 32GB Mac mini reports that it works — but “everything runs slowly.” That’s the honest baseline at the low end of viable hardware: functional, not fast.

The 2-bit quant trap

At this point someone always says: what about the community quants?

They exist. Unsloth and others publish 2-bit GGUFs that shrink Glimmer to 10.7–12.4GB — and yes, the smallest of those could squeeze under a 16GB Mac’s ~12.1GB Metal budget.

Here’s the trap: 2-bit quantization of a dense 30B model typically degrades tool-calling precision. And tool-calling is the reason to run this model. Remember the benchmark picture: independent testing found Glimmer barely edges Qwen3.6-27B outside of tool-calling, and actually loses on TerminalBench (51.7 vs 60.7). The community’s one-line summary is fair: best-in-class local tool-caller, well below hosted frontier quality and speed.

So a 2-bit Glimmer is a model whose distinctive strength has been sanded off, running at the edge of your memory budget, that was only middling outside that strength to begin with. You’ve made it fit by removing the reason you wanted it.

Can’t run it locally? Your hosted options

The obvious next thought — “fine, I’ll run it on Ollama Cloud” — doesn’t work either. Muse Glimmer is not available on Ollama Cloud, at any tier. Ollama Cloud’s Free, Pro ($20/mo), and Max ($100/mo) plans don’t include it, because it isn’t in the catalog at all.

Hosted access exists only through third parties — OpenRouter, Together, and Fireworks all serve it per-token. If you want to evaluate whether Glimmer’s tool-calling fits your workload before spending anything on hardware, that’s the sane route: pay per-token, test the model, then decide whether local inference earns its keep.

The honest math: buy outcomes, not hardware

Let’s put the two paths side by side, because we live this trade-off daily.

We run an autonomous agent in production on hosted frontier models — a mix of Anthropic and OpenAI. The entire LLM bill is about $2 a day, roughly $60–70 a month. For that we get frontier-quality reasoning at hosted speeds, no memory math, no quant compromises.

The local path for Glimmer: $2,000–4,000 of hardware to run a 30B model at 10–14 tokens per second — a model that, outside tool-calling, roughly ties a 27B competitor and sits well below hosted frontier quality. At $60–70/month of API spend, that hardware outlay buys years of frontier access before it breaks even, and the thing it buys you is a slower, weaker model.

Local inference has real, legitimate use cases — air-gapped environments, hard privacy constraints, tinkering for its own sake. If that’s you, buy the 32GB-minimum Mac with eyes open and enjoy it.

But if what you actually want is work done — leads followed up, emails handled, an agent that runs your business’s busywork — the hardware question is a distraction. That’s the premise behind MeetRick: an AI revenue agent that runs businesses’ AI for them, from $29/month, or fully managed at $499 — so the only spec you have to check is your to-do list.

Bottom line: Muse Glimmer is a real achievement and a genuinely good local tool-caller. But “can I run it?” has a numeric answer, and for the 8GB and 16GB Macs most people own, that answer is no — by about six gigabytes.