How Much RAM Does Ollama Need on a Mac? Real Tag Sizes by RAM Tier

Ollama does not publish a RAM requirement for macOS. Its macOS doc names an operating system and a chip family and stops there, so the number that decides whether a model runs on your Mac is the size of the tag you pull. That size is printed on each model's ollama.com page: 18GB for qwen3.8:27b, 19GB - 20GB for gemma4:31b, 105GB for the qwen3.8-flash-next MLX tag. This page lists the sizes as Ollama showed them on 2026-11-11 and maps them to 8GB, 16GB, 32GB, 64GB and 128GB Macs.

TL;DR: Ollama's own requirement is macOS Sonoma (v14) or newer on an Apple M series chip, with no RAM figure. RAM depends on the tag. Judged against a two-thirds-of-RAM yardstick, small tags (lfm3, gemma4:e2b) suit an 8GB Mac, tags up to the 8.7GB class suit 16GB, the 18GB - 20GB class (qwen3.8:27b, gemma4:31b) wants 32GB, the 39GB - 46GB ripple-preview tag is tight on 64GB, and the 105GB flash-next MLX tag only has a chance on 128GB. A tag's file size is a floor, not the footprint: context and macOS come on top.

What does Ollama itself require on a Mac?

The macOS doc in the Ollama repository lists two system requirements:

• macOS Sonoma (v14) or newer.
• Apple M series (CPU and GPU support) or x86 (CPU only).

There is no memory line. The same doc says downloaded models "can be tens to hundreds of GB in size", which is disk space, not RAM.

Since Ollama v0.40.0, model architectures that the MLX runtime supports run on MLX by default on Apple Silicon, as the release notes state. Those notes name no macOS version and no chip generation for the MLX path. Whether it needs anything newer than Sonoma is not verified, so this page repeats only the Sonoma line from the doc. What the default changes in practice is covered in our v0.40.0 write-up; the runtime speed question is in MLX vs Ollama on Mac. For first install and commands, see the Ollama install guide for Mac.

This is not a launch. Most of the tags below are weeks old: the gemma4 tags page shows the family as updated 4 weeks ago, and the qwen3.8 page lists the plain 27b tag at 1 month. The mimo-v2.6 and lfm3 pages show 1 week. Only ripple-preview and the qwen3.8-flash-next MLX tags show as updated yesterday.

How big is each tag, and which Mac fits it?

The sizes below are the labels Ollama prints, in GB. Ollama states GB, and this page does not convert them. A range such as 19GB - 20GB is Ollama's own wording for a tag that resolves to more than one build.

Yardstick for the Yes, Tight and No columns, which is ours and not Ollama's: the 16GB Mac guide counts about 11GB of a 16GB Mac as available for a model once macOS and apps take their share. That is roughly two thirds of the installed RAM. The table applies the same two thirds to every tier (about 5GB, 11GB, 21GB, 43GB and 85GB for 8GB, 16GB, 32GB, 64GB and 128GB), judged on the larger end of any range.

• Yes: the tag is at or under two thirds of the RAM, which leaves room for context.
• Tight: the tag is above two thirds but under the full RAM. It may load; short context only, and no other heavy apps.
• No: the tag is as large as the RAM or larger.

"Yes" is a size check against a yardstick, not a measurement: ModelFit did not run these tags for this page. The two-thirds rule comes from one sourced tier, 16GB. It is likely conservative on bigger Macs, where the OS takes a smaller share, but no source in hand gives a usable fraction for 32GB, 64GB or 128GB, so that part is not verified.

Mixture-of-experts tags do not save memory. The a4b and a6b in gemma4:26b-a4b and qwen3.8-flash-next:125b-a6b-nvfp4 name the active-parameter count, and the file you pull still holds every weight, which is why the flash-next tag is 105GB. Our 16GB Mac guide makes the same point for a 122B-A10B model: the speed behaves like a smaller model, the memory does not.

Why does Ollama show a range such as 19GB - 20GB?

Because the default tag points at more than one build. Comparing the ends of each range with the individual tags on the gemma4, mimo-v2.6, ripple-preview, qwen3.8 and qwen3.8-flash-next tag lists, the two ends match a q4_K_M tag and an nvfp4 MLX tag:

The MLX build is not always the smaller one. For gemma4 it is slightly smaller, for mimo-v2.6 and ripple-preview it is the larger one. On a Mac close to the limit, the range is the real uncertainty: plan on the larger end. Which build a given Mac receives from a plain ollama pull is not stated on the library pages, so check it after the pull: ollama list prints the SIZE of what landed on disk. For qwen3.8-flash-next the two builds differ the other way round: the nvfp4 MLX tag is 105GB and the q4_K_M tag is 120GB, so on a 128GB Mac the MLX tag is the one with any slack.

Is the tag size the RAM Ollama uses?

No. The tag size is the weights file. Three things sit on top of it, and Ollama's pages give no figure for any of them, so none is invented here:

• Context. The library pages list a context window per tag: 256K for qwen3.8 and the gemma4 26b and 31b tags, 2M for ripple-preview. That is the maximum. Memory for the context cache grows with the context you actually use, and a long agent session costs more than a short chat. No memory figure for it appears on the pages.
• macOS and your apps. The OS, the browser and the editor are resident before the model loads. This is the share the yardstick above tries to price in.
• The loaded model. ollama ps prints the size of the model as loaded. Read it on your own machine after a real prompt; it is the closest honest number for your setup.

The VRAM guide says a 16GB Mac behaves like a 12GB GPU once the OS has its share, and the 16GB Mac guide puts the pool at about 11GB. The two round the same overhead differently. A 14B model at Q4 loads on 16GB and then swaps, because a file that fits is not a model that runs.

What should you pull on each Mac?

8GB. Stay with the small tags: lfm3 at 1.1GB - 1.7GB, gemma4:e2b at 4.6GB. The e4b and mimo-v2.6 q4_K_M tags at 6.6GB are Tight, and the 12b class at 7.7GB - 8.0GB is the whole machine. The lfm3 page itself says that at 4-bit precision the model takes about 1 GB of memory.

16GB. The 7.7GB - 8.0GB gemma4:12b and the 8.7GB mimo-v2.6 nvfp4 tag pass the yardstick. The 16GB - 19GB gemma4:26b does not, and neither do the 18GB qwen3.8:27b tags. Best LLM for a 16GB Mac ranks what fits with engine numbers.

32GB. This is where the 27B-31B class starts: qwen3.8:27b at 18GB and gemma4:31b at 19GB - 20GB pass, with the larger context left to you. The q8_0 builds at 30GB and 34GB do not. For the Gemma 4 sizes in detail, see Gemma 4 GPU requirements; that guide sizes the same family for GPUs, and the tag sizes above are the Mac-side numbers.

64GB. The q8_0 builds of qwen3.8:27b (30GB) and gemma4:31b (34GB) fit. ripple-preview:65b is Tight: its 39GB q4_K_M tag passes the yardstick, its 46GB nvfp4 MLX tag does not. ripple-preview is a preview.

128GB. ripple-preview:65b passes and so does the 56GB bf16 build of qwen3.8:27b. The 105GB qwen3.8-flash-next MLX tag is Tight and the 120GB q4_K_M tag is Tight with almost nothing left, so treat both as an experiment. The library page describes it as an experimental preview of the architecture that will underpin Qwen4.

How to check the size on your own Mac

• ollama list shows the SIZE of every model on disk after a pull.
• ollama ps shows what is loaded right now, with its size and the processor split.
• Re-read the tag size on the model's ollama.com tags page before a big pull. These labels move when a tag is re-pushed, as the flash-next MLX tags were.

FAQ

How much RAM does Ollama need on a Mac?

Ollama states none. Its macOS doc lists Sonoma (v14) or newer and an Apple M series chip, and nothing about memory. The RAM you need is set by the tag: 1.1GB - 1.7GB for lfm3, 18GB for qwen3.8:27b, 105GB for the qwen3.8-flash-next MLX tag, plus room for context and macOS.

Can an 8GB Mac run Ollama?

Yes, with small tags. lfm3 at 1.1GB - 1.7GB and gemma4:e2b at 4.6GB pass the yardstick on 8GB. A tag near the full 8GB, such as gemma4:12b at 7.7GB - 8.0GB, is not a fit.

Does the MLX default change how much RAM a model needs?

Ollama does not document a memory difference. The tag lists show that the file size can go either way: the MLX nvfp4 build of gemma4:31b is 19GB against 20GB for q4_K_M, while for mimo-v2.6:9b it is 8.7GB against 6.6GB. Compare the exact tag sizes, not the runtime.

Can a 32GB Mac run qwen3.8:27b?

The 18GB tag is well under 32GB, so by the yardstick on this page it fits with room for context. We did not measure it. The 16GB Mac guide shows the same logic one tier down, where a file that fits can still swap.

Why does the same model show two different sizes on ollama.com?

The default tag can point at a GGUF build and an MLX build, and Ollama prints a range. gemma4:31b shows 19GB - 20GB because its q4_K_M tag is 20GB and its nvfp4 MLX tag is 19GB.

Sources

• Ollama macOS documentation (raw)
• Ollama v0.40.0 release notes
• qwen3.8 library page and qwen3.8 tags
• gemma4 tags
• mimo-v2.6 tags
• lfm3
• ripple-preview tags
• qwen3.8-flash-next tags