Guides
How much RAM do you need for local AI?
Memory decides which AI model your device can run, and the chip decides how fast. In Loci's measurements, phones with 4 GB run the smallest models, phones with 6 to 8 GB run 2B models such as Gemma 4 E2B, and Macs go from 2B models on 8 GB to 12B models on 24 GB.
What changed. New guide. It shows what phones and Macs with each memory size run in Loci, how fast, and the memory each model needs on a Mac.
Memory decides which AI model your device can run. The chip decides how fast it runs. A model has to be loaded into memory before it can answer, so the first question is what fits. This guide answers it with measurements: what phones and Macs with each amount of memory run in Loci, and how fast. We make Loci, and every speed here comes from its published data.
What phones run, by memory
Each row groups the phone models measured with that much memory.
iPhone
| Memory | Most used model | Median speed | Measured on |
|---|---|---|---|
| 4 GB | Qwen3.5 2B (2), LFM2.5 1.2B (1) | 9.3 to 16 tok/s | iPhone 12, iPhone 13, iPhone 11 |
| 6 GB | Gemma 4 E2B (2), Qwen3.5 2B (1) | 19 to 34 tok/s | iPhone 15, iPhone 14 Pro Max, iPhone 14 |
| 8 GB | Gemma 4 E2B | 38 to 54 tok/s | 6 iPhones, from the iPhone 17 to the iPhone 15 Pro Max |
| 12 GB | Gemma 4 E2B (1), Gemma 4 E4B (1) | 27 to 66 tok/s | iPhone 17 Pro, iPhone 17 Pro Max |
Android
| Memory | Most used model | Median speed | Measured on |
|---|---|---|---|
| 4 GB | LFM2.5 1.2B | 3.8 to 11 tok/s | 10 phones, from the moto g35 5G to the Galaxy A06 |
| 8 GB | Gemma 4 E2B | 5.7 to 17 tok/s | 5 phones, from the Galaxy S24 FE to the Pixel 8 |
| 12 GB | Gemma 4 E2B | 8.9 to 28 tok/s | 7 phones, from the Galaxy S25+ to the Pixel 9 |
Measured in Loci from October 10, 2026 to November 9, 2026, with Loci's default settings. Each phone model counts once, on the model most people run on it, from at least 10 phones. A number in brackets is how many phone models run that model most. Phone models measured with more than one memory size are left out.
Two things stand out:
- More memory lets a phone run a larger model. It does not make a model faster. With the same 8 GB of memory and the same model, Gemma 4 E2B, the iPhone 17 writes 54 tokens per second and the Pixel 8 writes 5.7.
- A larger model is slower on the same device. Each device's page in the speed statistics lists every model measured on it.
Phones side by side
| Phone | Chip | Memory | Most used model | Median speed |
|---|---|---|---|---|
| iPhone 17 Pro | A19 Pro | 12 GB | Gemma 4 E2B | 66 tok/s |
| iPhone 16 Pro | A18 Pro | 8 GB | Gemma 4 E2B | 49 tok/s |
| iPhone 16 Pro Max | A18 Pro | 8 GB | Gemma 4 E2B | 43 tok/s |
| iPhone 13 | A15 Bionic | 4 GB | Qwen3.5 2B | 15 tok/s |
| Galaxy S25 Ultra | Snapdragon 8 Elite | 12 GB | Gemma 4 E2B | 17 tok/s |
| Galaxy S22 Ultra | Snapdragon 8 Gen 1 | 8–12 GB | Gemma 4 E2B | 9.3 tok/s |
| Galaxy A36 5G | Snapdragon 6 Gen 3 | 6–8 GB | Qwen3.5 2B | 10 tok/s |
Each phone links to its page in the speed statistics, where every model measured on it is listed. Memory is that of the phones measured, so a phone sold in two sizes shows a range.
What Macs run, by memory
| Mac | Model | Speed | Measured on |
|---|---|---|---|
| M2 Mac with 8 GB | Qwen3.5 2B | 44 tokens per second | 2 Macs |
| M1 Mac with 8 GB | Qwen3.5 2B | 27 tokens per second | 2 Macs |
| M3 Mac with 8 GB | Qwen3.5 2B | 23 tokens per second | 2 Macs |
| M5 Mac with 16 GB | Gemma 4 E4B | 33 tokens per second | 2 Macs |
| M4 Mac with 16 GB | Gemma 4 E4B | 21 tokens per second | 3 Macs |
| M5 Pro Mac with 24 GB | Gemma 4 12B | 31 tokens per second | 6 Macs |
| M4 Pro Mac with 24 GB | Gemma 4 12B | 25 tokens per second | 2 Macs |
| M5 Mac with 24 GB | Gemma 4 12B | 15 tokens per second | 2 Macs |
| M2 Pro Mac with 32 GB | Gemma 4 12B | 18 tokens per second | 2 Macs |
| M5 Pro Mac with 48 GB | Gemma 4 26B A4B | 45 tokens per second | 2 Macs |
| M5 Max Mac with 128 GB | Qwen3.8 27B Max Quality | 11 tokens per second | 3 Macs |
Measured in Loci from October 10, 2026 to November 9, 2026; each row is the median across the Macs counted, with Loci's default settings. Each row comes from a small number of Macs, shown in the last column, so read it as an indication.
How much memory each model needs on a Mac
| Model | Download | Mac memory |
|---|---|---|
| LFM2.5 1.2B | 0.7 GB | 3 GB or more |
| Qwen3.5 2B | 1.3 GB | 4 GB or more |
| Bonsai 2 4B | 1.2 GB | 6 GB or more |
| Qwen3.5 4B | 2.7 GB | 6 GB or more |
| Gemma 4 E2B | 3.1 GB | 6 GB or more |
| Gemma 4 E4B | 5.0 GB | 12 GB or more |
| Qwen3.5 9B | 5.7 GB | 16 GB or more |
| Gemma 4 12B | 7.1 GB | 16 GB or more |
| Ternary Bonsai 2 27B | 7.2 GB | 24 GB or more |
| Qwen3.8 27B Max Quality | 29 GB | 64 GB or more |
Memory is the least installed RAM the Loci Mac catalog lists for each model, not a promise of fit. Catalog checked October 30, 2026. With 8 GB, a Mac can run LFM2.5 1.2B, Qwen3.5 2B, Bonsai 2 4B, Qwen3.5 4B and Gemma 4 E2B.
Why memory is the limit
- The whole model sits in memory while it runs. Its download size is a good first estimate of the space it takes.
- The conversation takes memory too. Short chats need less than long documents, and the system and your other open apps need room as well.
- Storage is a separate budget. Storage holds the downloaded file, and memory runs it. Model downloads in Loci's Mac catalog range from 0.7 GB (LFM2.5 1.2B) to 29 GB (Qwen3.8 27B Max Quality).
- On a PC, a graphics card has its own memory. A model runs fastest when all of it fits there. The best local LLMs guide compares 16 GB of RAM with 16 GB of VRAM.
How to check your memory
- Mac: open the Apple menu and choose About This Mac.
- Windows: open Task Manager, choose Performance, and read Memory. With a separate graphics card, also read Dedicated GPU memory under GPU.
On a phone you do not need to look it up: Loci picks a model that fits the phone's memory.
What to read next
What is local AI? covers the basics, best local LLMs compares the model families, and the device checker tests one model against one phone or Mac. The comparison pages set the local AI apps side by side.
FAQ
How much RAM do I need for local AI?
Enough to hold the model you want to run. In Loci's measurements, LFM2.5 1.2B is the most used model on phones with 4 GB of memory (11 of 13 phone models), and Gemma 4 E2B on phones with 8 GB or more (19 of 20). On the Macs measured in Loci, the most used model is Qwen3.5 2B on 8 GB, Gemma 4 E4B on 16 GB, Gemma 4 12B on 24 GB, Gemma 4 12B on 32 GB, Gemma 4 26B A4B on 48 GB and Qwen3.8 27B Max Quality on 128 GB.
Is 8 GB of RAM enough for local AI?
Yes, for small models. With 8 GB, a Mac can run LFM2.5 1.2B, Qwen3.5 2B, Bonsai 2 4B, Qwen3.5 4B and Gemma 4 E2B in Loci. The phone tables above show what 8 GB phones run.
Does more RAM make local AI faster?
Not by itself. Memory decides which models fit, and the chip sets the speed. With the same 8 GB of memory and the same model, Gemma 4 E2B, the iPhone 17 writes 54 tokens per second and the Pixel 8 writes 5.7.
What is the best local LLM for a phone?
There is no single best one. In Loci's measurements, LFM2.5 1.2B is the most used model on phones with 4 GB of memory (11 of 13 phone models), and Gemma 4 E2B on phones with 8 GB or more (19 of 20). Best local LLMs compares the model families.
Which model would be best for my device?
The largest model that fits your memory and still feels fast. Loci picks one for you, and the tables above show what people with the same memory run most.
How much RAM does a local LLM need on a Mac?
On the Macs measured in Loci, the most used model is Qwen3.5 2B on 8 GB, Gemma 4 E4B on 16 GB, Gemma 4 12B on 24 GB, Gemma 4 12B on 32 GB, Gemma 4 26B A4B on 48 GB and Qwen3.8 27B Max Quality on 128 GB. The catalog table above lists the memory each model needs.
Is RAM the same as storage?
No. Storage holds the downloaded model file. RAM is where the model is loaded while it runs. You need enough of both.
Do I need a graphics card for local AI?
No. Small models run on a recent phone, an Apple silicon Mac or a Windows 11 laptop without one. A graphics card helps with larger models when the model fits in the card's memory.
Get Loci
Private AI that keeps working.
Download Loci once, then chat without an account, chat upload, or signal.