Catalog snapshot: November 15, 2026. Counts below are fit estimates, rather than a claim that every model has been tested on every Mac.
The short answer
- 8 GB: start with small chat models and short conversations.
- 16 GB: more room for chat, including Nemotron 3.5 Bolt E4B. Check the model size before downloading.
- 24 to 32 GB: consider these sizes if a model you want is too large for 16 GB, or you keep several apps open.
- 48 GB and up: compare the specific model and context length before paying for the extra memory.
What each memory size runs
Pick a size. Each answer is for one Mac at that size, named in the panel. Chips with the same RAM can have different graphics-memory limits. Use the linked Mac page for your exact configuration.
8 GB: 8 of 51 models run well
Worked out for a Mac with an M3 and 8 GB.
- Chat
- 8 of 29
- Pictures
- 0 of 13
- Video
- 0 of 9
| Model | Size to get | Memory used | Tokens a second |
|---|---|---|---|
| Gemma 4 E2B | Q4_K_M, 4.1 GB | 5.2 GB32,768 tokens of context | 20 est.Estimate at 4K context |
| Qwen3.5 4B | Q4_K_M, 3.4 GB | 5.2 GB32,768 tokens of context | 18 est.Estimate at 4K context |
| Qwen3.5 2B | Q6_K, 2.2 GB | 3.5 GB32,768 tokens of context | 33 est.Estimate at 4K context |
- Pictures: no picture model in the catalog runs well on 8 GB.
- Video: no video model in the catalog runs well on 8 GB.
16 GB: 20 of 51 models run well
Worked out for a Mac with an M4 and 16 GB.
- Chat
- 12 of 29
- Pictures
- 4 of 13
- Video
- 4 of 9
| Model | Size to get | Memory used | Tokens a second |
|---|---|---|---|
| Nemotron 3.5 Bolt E4B | UD-Q4_K_XL, 5.1 GB | 6.6 GB32,768 tokens of context | 15 est.Estimate at 4K context |
| K2-Horizon 7B | Q4_K_M, 5.8 GB | 7.1 GB8,192 tokens of context | 8.5 est.Estimate at 4K context |
| K2-Horizon 3.7B | Q6_K, 4.4 GB | 7.6 GB32,768 tokens of context | 9.8 est.Estimate at 4K context |
- Pictures: NitroFusion 2 runs well, using 9.0 GB.
- Video: Wan 2.2 14B Image to Video runs well, using 8.3 GB.
24 GB: 30 of 51 models run well
Worked out for a Mac with an M4 and 24 GB.
- Chat
- 18 of 29
- Pictures
- 6 of 13
- Video
- 6 of 9
| Model | Size to get | Memory used | Tokens a second |
|---|---|---|---|
| Qwen3.8 27B | UD-IQ4_XS, 15 GB | 16 GB8,192 tokens of context | 5.4 est.Estimate at 4K context |
| Nemotron 3.5 Bolt E4B | UD-Q6_K_XL, 7.3 GB | 9.0 GB32,768 tokens of context | 11 est.Estimate at 4K context |
| K2-Horizon 7B | Q6_K, 7.8 GB | 11 GB32,768 tokens of context | 7.2 est.Estimate at 4K context |
- Pictures: NitroFusion 2 runs well, using 9.0 GB.
- Video: LTX-2.3 runs well, using 15 GB.
32 GB: 38 of 51 models run well
Worked out for a Mac with an M4 and 32 GB.
- Chat
- 18 of 29
- Pictures
- 11 of 13
- Video
- 9 of 9
| Model | Size to get | Memory used | Tokens a second |
|---|---|---|---|
| Nemotron 3.5 Bolt E4B | UD-Q6_K_XL, 7.3 GB | 9.0 GB32,768 tokens of context | 11 est.Estimate at 4K context |
| Qwen3.8 27B | UD-Q4_K_XL, 18 GB | 20 GB32,768 tokens of context | 4.5 est.Estimate at 4K context |
| K2-Horizon 7B | Q6_K, 7.8 GB | 11 GB32,768 tokens of context | 7.2 est.Estimate at 4K context |
- Pictures: Qwen Image Edit 2511 runs well, using 21 GB.
- Video: LTX-2.3 runs well, using 15 GB.
48 GB: 39 of 51 models run well
Worked out for a Mac with an M4 Pro and 48 GB.
- Chat
- 18 of 29
- Pictures
- 12 of 13
- Video
- 9 of 9
| Model | Size to get | Memory used | Tokens a second |
|---|---|---|---|
| Gemma 4 26B | UD-Q4_K_XL, 18 GB | 22 GB131,072 tokens of context | 32 est.Estimate at 4K context |
| Qwen3.8 27B | UD-Q4_K_XL, 18 GB | 20 GB32,768 tokens of context | 9.6 est.Estimate at 4K context |
| Gemma 4 31B | Q4_K_M, 20 GB | 23 GB32,768 tokens of context | 9.1 est.Estimate at 4K context |
- Pictures: MiniMax-H3-Image runs well, using 26 GB.
- Video: LTX-2.3 runs well, using 19 GB.
64 GB: 41 of 51 models run well
Worked out for a Mac with an M4 Max and 64 GB.
- Chat
- 19 of 29
- Pictures
- 13 of 13
- Video
- 9 of 9
| Model | Size to get | Memory used | Tokens a second |
|---|---|---|---|
| Gemma 4 26B | Q8_0, 28 GB | 30 GB32,768 tokens of context | 47 est.Estimate at 4K context |
| Qwen3.8 27B | UD-Q6_K, 23 GB | 28 GB131,072 tokens of context | 15 est.Estimate at 4K context |
| Gemma 4 31B | Q6_K, 26 GB | 30 GB32,768 tokens of context | 13 est.Estimate at 4K context |
- Pictures: MiniMax-H3-Image runs well, using 26 GB.
- Video: LTX-2.3 runs well, using 24 GB.
128 GB: 45 of 51 models run well
Worked out for a Mac with an M4 Max and 128 GB.
- Chat
- 23 of 29
- Pictures
- 13 of 13
- Video
- 9 of 9
| Model | Size to get | Memory used | Tokens a second |
|---|---|---|---|
| Gemma 4 26B | Q8_0, 28 GB | 34 GB262,144 tokens of context | 47 est.Estimate at 4K context |
| Qwen3.8 27B | UD-Q6_K, 23 GB | 34 GB262,144 tokens of context | 15 est.Estimate at 4K context |
| Gemma 4 31B | Q6_K, 26 GB | 43 GB262,144 tokens of context | 13 est.Estimate at 4K context |
- Pictures: MiniMax-H3-Image runs well, using 43 GB.
- Video: LTX-2.3 runs well, using 24 GB.
How the count grows
Models that run well, of the 51 in Outmake’s catalog, at each memory size.
Why memory decides it
On an Apple silicon Mac the processor and the graphics chip share one pool of memory. The model’s weights and working memory share that pool with macOS and your apps. Outmake can load some picture and video models in parts. Those models can use less memory than their download size suggests, at the cost of speed.
A model’s size comes from two numbers: how many parameters it has, and how many bits each is stored in. A 4-bit version uses fewer bits per weight than a 16-bit version. The complete download and working memory also include other files and overhead. Lower precision can reduce answer quality; the effect varies by model and task. For example, Nemotron 3.5 Bolt E4B at UD-Q4_K_XL is a 5.1 GB download and uses 6.6 GB at 32,768 tokens of context. The chat so far takes memory too, so a long conversation or a long document needs more than a short one.
Outmake gives each model one of four labels on your Mac: Runs great, Runs well, Tight fit or Too big. This page counts the first two, which leave room for the rest of the Mac. Tight fit means quitting other large apps first.
Why chip choice matters
Memory capacity determines which model sizes are practical. Memory bandwidth, the inference engine and the task also affect speed. The figures below are the bandwidth values used by Outmake’s estimates.
| Mac | Memory | Memory speed |
|---|---|---|
| M3 | 8 GB | 100 GB a second |
| M4 | 16 GB | 120 GB a second |
| M4 | 24 GB | 120 GB a second |
| M4 | 32 GB | 120 GB a second |
| M4 Pro | 48 GB | 273 GB a second |
| M4 Max | 64 GB | 546 GB a second |
| M4 Max | 128 GB | 546 GB a second |
A higher bandwidth can help text generation, but it does not guarantee a proportional speed increase. Estimated speeds use that figure and the model’s size, and this page says how they’re made and how wrong they can be.
If you’re buying a Mac
- Choose enough memory for the models you need now, with room for your other apps. Check Apple’s specifications for the Mac you plan to buy.
- For a model that is too large for your current Mac, extra memory may matter more than a newer chip. For a model that fits, compare measured speed as well.
- Budget for storage too. Check the complete model download, including companion files. Outmake can keep models on an external drive; slow drives can make loading and generation take longer.
Check your own Mac
Find your Mac to see every model it runs, the size to download and how fast it writes. In the app, each model in Models carries its label for the Mac you’re on, worked out from its memory and what is running.
How this page is made
The counts and picks are worked out each morning from Outmake’s catalog and the same fit judgement the app uses. Each model is counted once, using the size the store selects for that Mac. Counts include selected low-bit versions. The three chat examples are filtered to 4-bit names or higher and follow the store’s ranking, which combines fit and available benchmark ratings. These are examples from the curated catalog, not a ranking of every model on Hugging Face. See the ranking method. Chat memory is shown at the context length selected by the store. Estimated writing speeds use 4,096 tokens of context. Image and video memory depend on the model’s documented settings. A fit label does not promise fast output or support for every feature of a model.
Sources and limits
- Outmake’s model data and Mac configurations, refreshed with the catalog.
- Memory and speed methodology, with measured results, sample counts and limitations.
- Apple’s Mac mini specifications and llama.cpp, the chat engine used by Outmake.
Published 11 November 2026. The page’s examples update with the catalog. No Mac was purchased or benchmarked for this guide.