Outlier · Local AI for Mac

What AI models can a Mac with 8 GB run?

Quick answer
  • A Mac with 8 GB fits 5 models for short chats, 4 for long 32,000-token work, from 33 models in our list. We use measured minimums where available and our cautious rule for the rest.
  • For long work, the roomiest fit is Qwen3.5-4B. It has the lowest listed memory floor, 6 GB, leaving 2 GB above that floor.
  • For long work, the smallest fit by parameter count is Qwen2.5-0.5B-Instruct, listed as 0.5B dense. It takes 0.28 GB on disk. The largest fit by parameter count is Qwen3.5-4B, listed as 4B dense.

Every model that fits

Fit, memory left and upgrade advice below use the long-work minimum. The two minimum columns show what changes for short chats. Measured rows use their measured minimum in both columns.

Start with the bottom of the hardware table. This memory tier includes older entry-level Macs, so the useful question is which downloads clear a small memory budget. Read the fit column before looking at speed. A file that is small enough to store can still ask for more working memory than this configuration has. The answer box identifies the model with the lowest floor and shows how much room remains above it.

For a first choice here, compare the smallest parameter counts with their separate memory minimums. The list is short enough to check each model. Some tiny models appear in the excluded table because our rule reserves more room for them. Their names alone cannot tell you whether they belong on this Mac. Follow the next-memory column to see the actual jump needed for each candidate.

The speed samples below cover only models in this tier. Pick your exact chip row to read the source behind a range. If your aim is to try several task types, first check which fitting models carry those tasks in the calculator. Keep the working conversation within the stated context when assessing whether this entry point meets your needs.

This is the lowest memory size listed for a Mac in our Mac list. Its complete fit list is Qwen2.5-0.5B-Instruct, Qwen3.5-0.8B, Llama-3.2-1B-Instruct, Qwen3.5-4B. Read the minimum beside each name before choosing. A smaller weight file can still have a higher memory floor in these rules.

Models that fit long work, with measured minimums used where available
ModelFitShort chatsLong 32,000-token work Memory left for long workDisk
Qwen2.5-0.5B-Instruct tight 8 GB 8 GB 0 GB 0.28 GB
Qwen3.5-0.8B tight 8 GB 8 GB 0 GB 0.63 GB
Llama-3.2-1B-Instruct tight 8 GB 8 GB 0 GB 0.7 GB
Qwen3.5-4B fits 6 GB measured 6 GB measured 2 GB 2.37 GB

Speed table by chip

The table covers all 11 Mac models sold with 8 GB in our Mac list, including older models. Each row names the exact Mac and chip. The representative models are Qwen3.5-4B, Qwen2.5-0.5B-Instruct, Llama-3.2-1B-Instruct, Qwen3.5-0.8B. They cover the roomiest fit, the smallest model, the largest model and the middle of this fit list, with repeated names removed.

Speed by chip for every Mac listed with 8 GB
Mac and chipQwen3.5-4BQwen2.5-0.5B-InstructLlama-3.2-1B-InstructQwen3.5-0.8B
MacBook Air M1Apple M1 about 4.9 to 14 tokens/s estimate, not measured by us
How we estimate speed 4.89 to 13.56 tokens/s; estimate: low end from memory bandwidth (Mac bandwidth 68.25 GB/s against 800 GB/s), high end from this chip's speed in llama.cpp's public Apple-silicon benchmark; measured on an M1 Ultra 64 GB: 71.7 tokens/s at 800 GB/s
about 49 to 98 tokens/s estimate, not measured by us
How we estimate speed 48.75 to 97.5 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 68.25 GB/s divided by active weights in GB
about 20 to 39 tokens/s estimate, not measured by us
How we estimate speed 19.5 to 39 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 68.25 GB/s divided by active weights in GB
about 22 to 43 tokens/s estimate, not measured by us
How we estimate speed 21.66 to 43.34 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 68.25 GB/s divided by active weights in GB
MacBook Pro 13 M1Apple M1 about 4.9 to 14 tokens/s estimate, not measured by us
How we estimate speed 4.89 to 13.56 tokens/s; estimate: low end from memory bandwidth (Mac bandwidth 68.25 GB/s against 800 GB/s), high end from this chip's speed in llama.cpp's public Apple-silicon benchmark; measured on an M1 Ultra 64 GB: 71.7 tokens/s at 800 GB/s
about 49 to 98 tokens/s estimate, not measured by us
How we estimate speed 48.75 to 97.5 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 68.25 GB/s divided by active weights in GB
about 20 to 39 tokens/s estimate, not measured by us
How we estimate speed 19.5 to 39 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 68.25 GB/s divided by active weights in GB
about 22 to 43 tokens/s estimate, not measured by us
How we estimate speed 21.66 to 43.34 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 68.25 GB/s divided by active weights in GB
Mac mini M1Apple M1 about 4.9 to 14 tokens/s estimate, not measured by us
How we estimate speed 4.89 to 13.56 tokens/s; estimate: low end from memory bandwidth (Mac bandwidth 68.25 GB/s against 800 GB/s), high end from this chip's speed in llama.cpp's public Apple-silicon benchmark; measured on an M1 Ultra 64 GB: 71.7 tokens/s at 800 GB/s
about 49 to 98 tokens/s estimate, not measured by us
How we estimate speed 48.75 to 97.5 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 68.25 GB/s divided by active weights in GB
about 20 to 39 tokens/s estimate, not measured by us
How we estimate speed 19.5 to 39 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 68.25 GB/s divided by active weights in GB
about 22 to 43 tokens/s estimate, not measured by us
How we estimate speed 21.66 to 43.34 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 68.25 GB/s divided by active weights in GB
iMac M1Apple M1 about 4.9 to 14 tokens/s estimate, not measured by us
How we estimate speed 4.89 to 13.56 tokens/s; estimate: low end from memory bandwidth (Mac bandwidth 68.25 GB/s against 800 GB/s), high end from this chip's speed in llama.cpp's public Apple-silicon benchmark; measured on an M1 Ultra 64 GB: 71.7 tokens/s at 800 GB/s
about 49 to 98 tokens/s estimate, not measured by us
How we estimate speed 48.75 to 97.5 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 68.25 GB/s divided by active weights in GB
about 20 to 39 tokens/s estimate, not measured by us
How we estimate speed 19.5 to 39 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 68.25 GB/s divided by active weights in GB
about 22 to 43 tokens/s estimate, not measured by us
How we estimate speed 21.66 to 43.34 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 68.25 GB/s divided by active weights in GB
MacBook Air M2Apple M2 about 7.2 to 21 tokens/s estimate, not measured by us
How we estimate speed 7.17 to 20.94 tokens/s; estimate: low end from memory bandwidth (Mac bandwidth 100 GB/s against 800 GB/s), high end from this chip's speed in llama.cpp's public Apple-silicon benchmark; measured on an M1 Ultra 64 GB: 71.7 tokens/s at 800 GB/s
about 71 to 143 tokens/s estimate, not measured by us
How we estimate speed 71.42 to 142.86 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
about 29 to 57 tokens/s estimate, not measured by us
How we estimate speed 28.57 to 57.15 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
about 32 to 63 tokens/s estimate, not measured by us
How we estimate speed 31.74 to 63.5 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
MacBook Pro 13 M2Apple M2 about 7.2 to 21 tokens/s estimate, not measured by us
How we estimate speed 7.17 to 20.94 tokens/s; estimate: low end from memory bandwidth (Mac bandwidth 100 GB/s against 800 GB/s), high end from this chip's speed in llama.cpp's public Apple-silicon benchmark; measured on an M1 Ultra 64 GB: 71.7 tokens/s at 800 GB/s
about 71 to 143 tokens/s estimate, not measured by us
How we estimate speed 71.42 to 142.86 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
about 29 to 57 tokens/s estimate, not measured by us
How we estimate speed 28.57 to 57.15 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
about 32 to 63 tokens/s estimate, not measured by us
How we estimate speed 31.74 to 63.5 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
Mac mini M2Apple M2 about 7.2 to 21 tokens/s estimate, not measured by us
How we estimate speed 7.17 to 20.94 tokens/s; estimate: low end from memory bandwidth (Mac bandwidth 100 GB/s against 800 GB/s), high end from this chip's speed in llama.cpp's public Apple-silicon benchmark; measured on an M1 Ultra 64 GB: 71.7 tokens/s at 800 GB/s
about 71 to 143 tokens/s estimate, not measured by us
How we estimate speed 71.42 to 142.86 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
about 29 to 57 tokens/s estimate, not measured by us
How we estimate speed 28.57 to 57.15 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
about 32 to 63 tokens/s estimate, not measured by us
How we estimate speed 31.74 to 63.5 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
MacBook Air 13 M3Apple M3 about 7.2 to 20 tokens/s estimate, not measured by us
How we estimate speed 7.17 to 20.44 tokens/s; estimate: low end from memory bandwidth (Mac bandwidth 100 GB/s against 800 GB/s), high end from this chip's speed in llama.cpp's public Apple-silicon benchmark; measured on an M1 Ultra 64 GB: 71.7 tokens/s at 800 GB/s
about 71 to 143 tokens/s estimate, not measured by us
How we estimate speed 71.42 to 142.86 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
about 29 to 57 tokens/s estimate, not measured by us
How we estimate speed 28.57 to 57.15 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
about 32 to 63 tokens/s estimate, not measured by us
How we estimate speed 31.74 to 63.5 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
MacBook Air 15 M3Apple M3 about 7.2 to 20 tokens/s estimate, not measured by us
How we estimate speed 7.17 to 20.44 tokens/s; estimate: low end from memory bandwidth (Mac bandwidth 100 GB/s against 800 GB/s), high end from this chip's speed in llama.cpp's public Apple-silicon benchmark; measured on an M1 Ultra 64 GB: 71.7 tokens/s at 800 GB/s
about 71 to 143 tokens/s estimate, not measured by us
How we estimate speed 71.42 to 142.86 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
about 29 to 57 tokens/s estimate, not measured by us
How we estimate speed 28.57 to 57.15 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
about 32 to 63 tokens/s estimate, not measured by us
How we estimate speed 31.74 to 63.5 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
iMac M3Apple M3 about 7.2 to 20 tokens/s estimate, not measured by us
How we estimate speed 7.17 to 20.44 tokens/s; estimate: low end from memory bandwidth (Mac bandwidth 100 GB/s against 800 GB/s), high end from this chip's speed in llama.cpp's public Apple-silicon benchmark; measured on an M1 Ultra 64 GB: 71.7 tokens/s at 800 GB/s
about 71 to 143 tokens/s estimate, not measured by us
How we estimate speed 71.42 to 142.86 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
about 29 to 57 tokens/s estimate, not measured by us
How we estimate speed 28.57 to 57.15 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
about 32 to 63 tokens/s estimate, not measured by us
How we estimate speed 31.74 to 63.5 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 100 GB/s divided by active weights in GB
MacBook NeoApple A18 Pro about 4.3 to 5.9 tokens/s estimate, not measured by us
How we estimate speed 4.3 to 5.92 tokens/s; estimate from memory bandwidth; measured on an M1 Ultra 64 GB: 71.7 tokens/s at 800 GB/s; Mac bandwidth 60 GB/s
about 43 to 86 tokens/s estimate, not measured by us
How we estimate speed 42.85 to 85.72 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 60 GB/s divided by active weights in GB
about 17 to 34 tokens/s estimate, not measured by us
How we estimate speed 17.14 to 34.29 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 60 GB/s divided by active weights in GB
about 19 to 38 tokens/s estimate, not measured by us
How we estimate speed 19.04 to 38.1 tokens/s; estimate, not measured by us; 0.20 to 0.40 x Mac bandwidth 60 GB/s divided by active weights in GB

What does not fit, and the next memory size

29 models exceed 8 GB for long work.

The next listed memory size is 16 GB. It admits 6 more models from this table. Those models are Qwen3-0.6B, Qwen3-4B-Instruct-2507, gemma-3-4b-it-qat, Qwen2.5-Coder-7B-Instruct, Qwen3.5-9B, Gemma-4-26B-a4b-it. The last column below gives the next suitable size for each excluded model, including sizes outside these memory guides.

Models needing more memory for long work, with their short-chat minimums alongside
ModelShort chatsLong 32,000-token work Memory short for long workNext listed memory
Qwen3-0.6B 8 GB 16 GB 8 GB 16 GB
Qwen3-4B-Instruct-2507 16 GB 16 GB 8 GB 16 GB
gemma-3-4b-it-qat 16 GB 16 GB 8 GB 16 GB
Mistral-7B-Instruct-v0.3 16 GB 18 GB 10 GB 18 GB
Qwen2.5-Coder-7B-Instruct 16 GB 16 GB 8 GB 16 GB
Meta-Llama-3.1-8B-Instruct 16 GB 18 GB 10 GB 18 GB
Qwen3-8B 16 GB 24 GB 16 GB 24 GB
Qwen3.5-9B 12 GB measured 12 GB measured 4 GB 16 GB
gemma-3-12b-it-qat 24 GB 48 GB 40 GB 48 GB
phi-4 24 GB 32 GB 24 GB 32 GB
Qwen3-14B 24 GB 32 GB 24 GB 32 GB
gpt-oss-20b 24 GB 24 GB 16 GB 24 GB
Mistral-Small-3.2-24B-Instruct-2506 32 GB 36 GB 28 GB 36 GB
Devstral-Small-2-24B-Instruct-2512 32 GB 36 GB 28 GB 36 GB
Gemma-4-26B-a4b-it 16 GB measured 16 GB measured 8 GB 16 GB
Qwen3.6-27B 24 GB measured 24 GB measured 16 GB 24 GB
Qwen3.8-27B 24 GB measured 24 GB measured 16 GB 24 GB
gemma-3-27b-it-qat 48 GB 64 GB 56 GB 64 GB
Qwen3-30B-A3B 32 GB 36 GB 28 GB 36 GB
Qwen3-Coder-30B-A3B-Instruct 32 GB 36 GB 28 GB 36 GB
gemma-4-31b-it 48 GB 96 GB 88 GB 96 GB
DeepSeek-R1-Distill-Qwen-32B 36 GB 48 GB 40 GB 48 GB
Qwen3-32B 36 GB 48 GB 40 GB 48 GB
Qwen3.6-35B-A3B 48 GB 48 GB 40 GB 48 GB
Llama-3.3-70B-Instruct 64 GB 96 GB 88 GB 96 GB
Llama-4-Scout-17B-16E-Instruct 96 GB 128 GB 120 GB 128 GB
gpt-oss-120b 96 GB 96 GB 88 GB 96 GB
Qwen3-235B-A22B 192 GB 256 GB 248 GB 256 GB
Qwen3.5-397B-A17B 64 GB measured 64 GB measured 56 GB 64 GB

Context length and data dates

Memory minimums use a 32K context; longer contexts need more memory. Default context is 32768 tokens.

1 fitting model has a M1 Ultra measurement. 3 models have rule-based minimums without our measurements. Figures as of 8 November 2026.

Mac source.

Apple technical specifications (apple.com/mac and the specs page of each Mac, read 2026-11-07; the iMac M6 page says only "up to 170GB/s", so its low end, 153 GB/s, is the 16 GB figure for the same M6 chip on the Mac mini page); table last changed 2026-11-07

Chip ratio source.

llama.cpp's public Apple-silicon benchmark (github.com/ggml-org/llama.cpp/discussions/4167, read 2026-11-07): LLaMA 7B Q4_0 text generation, as a ratio to the M1 Ultra 48-core GPU

Model source.

Dates come from each source text; otherwise table last changed 2026-10-19

Reference test machine.

M1 Ultra, 64 GB, 800 GB/s
Source date and evidence for every model discussed
ModelSource dateEvidence
Qwen3.5-4B 2026-04-29 Measured on M1 Ultra, measured on 2026-04-29; the build details weren't recorded with this figure
Qwen3.5-9B table last changed 2026-10-19 Measured on M1 Ultra, the build and date weren't recorded with this figure
Gemma-4-26B-a4b-it 2026-08-24 Measured on M1 Ultra, from build 1.11.804 on 2026-08-24
Qwen3.6-27B table last changed 2026-10-19 Measured on M1 Ultra, the build and date weren't recorded with this figure
Qwen3.5-397B-A17B 2026-05-01 Measured on M1 Ultra, measured on 2026-05-01; the build details weren't recorded with this figure
Qwen3.8-27B 2026-08-19 Measured on M1 Ultra, from build 1.11.792 on 2026-08-19
Llama-3.3-70B-Instruct 2026-11-08 Rule-based minimum, not measured by us
Qwen2.5-Coder-7B-Instruct 2026-11-08 Rule-based minimum, not measured by us
Qwen3-4B-Instruct-2507 2026-11-08 Rule-based minimum, not measured by us
Devstral-Small-2-24B-Instruct-2512 2026-11-08 Rule-based minimum, not measured by us
Qwen3-0.6B 2026-11-08 Rule-based minimum, not measured by us
gemma-3-4b-it-qat 2026-11-08 Rule-based minimum, not measured by us
Qwen3.6-35B-A3B 2026-11-08 Rule-based minimum, not measured by us
Qwen3.5-0.8B 2026-11-08 Rule-based minimum, not measured by us
Qwen2.5-0.5B-Instruct 2026-11-08 Rule-based minimum, not measured by us
Llama-3.2-1B-Instruct 2026-11-08 Rule-based minimum, not measured by us
Meta-Llama-3.1-8B-Instruct 2026-11-08 Rule-based minimum, not measured by us
Mistral-7B-Instruct-v0.3 2026-11-08 Rule-based minimum, not measured by us
Mistral-Small-3.2-24B-Instruct-2506 2026-11-08 Rule-based minimum, not measured by us
gemma-3-12b-it-qat 2026-11-08 Rule-based minimum, not measured by us
gemma-3-27b-it-qat 2026-11-08 Rule-based minimum, not measured by us
gemma-4-31b-it 2026-11-08 Rule-based minimum, not measured by us
Qwen3-8B 2026-11-08 Rule-based minimum, not measured by us
Qwen3-14B 2026-11-08 Rule-based minimum, not measured by us
Qwen3-32B 2026-11-08 Rule-based minimum, not measured by us
Qwen3-Coder-30B-A3B-Instruct 2026-11-08 Rule-based minimum, not measured by us
Qwen3-30B-A3B 2026-11-08 Rule-based minimum, not measured by us
gpt-oss-20b 2026-11-08 Rule-based minimum, not measured by us
gpt-oss-120b 2026-11-08 Rule-based minimum, not measured by us
DeepSeek-R1-Distill-Qwen-32B 2026-11-08 Rule-based minimum, not measured by us
phi-4 2026-11-08 Rule-based minimum, not measured by us
Llama-4-Scout-17B-16E-Instruct 2026-11-08 Rule-based minimum, not measured by us
Qwen3-235B-A22B 2026-11-08 Rule-based minimum, not measured by us

Outlier Pro: $39 once

Download free Buy Pro

Free to start (Nano and Lite). macOS 15+, Apple silicon. Refund window: 30 days, no questions asked. Priced in US dollars. Checkout shows your currency and any local tax.

Try Outlier on your Mac

Use the memory table and the speed source labels above when choosing a local model.

Download Outlier