What can the Apple M3 run?
With 16 GB, 4 of 12 popular open models fit. The best pick is Qwen3 4B at about 22 tok/s, about reading speed.
ollama run qwen3:4bFirst time? Install Ollama, then paste that into a terminal. Is this your machine? Check it on the homepage.
Every model the Apple M3 8 GB can run
1 of 12 fit. Best pick: Qwen3 4B.
- Qwen3 4BQuick everyday chat and summaries~22 tok/sUsable
qwen3:4b - Qwen3 8BA solid all-rounder for chat and writingneeds 7 GBToo big
qwen3:8b - Gemma 3 12BReads images as well as textneeds 10 GBToo big
gemma3:12b - Qwen3 14BNoticeably smarter at reasoning and codeneeds 11 GBToo big
qwen3:14b - gpt-oss 20BOpenAI's open model. Fast, strong reasoningneeds 16 GBToo big
gpt-oss:20b - Gemma 3 27BGoogle's best open model for its sizeneeds 19 GBToo big
gemma3:27b - Qwen3 30B-A3B30B smarts at small-model speedneeds 21 GBToo big
qwen3:30b - Qwen3 32BStrong coding and long answersneeds 22 GBToo big
qwen3:32b - Llama 3.3 70BClose to the big cloud modelsneeds 45 GBToo big
llama3.3:70b - gpt-oss 120BOpenAI's largest open modelneeds 67 GBToo big
gpt-oss:120b - Qwen3 235B-A22BAlibaba's flagship open model, for multi-GPU rigsneeds 144 GBToo big
qwen3:235b - DeepSeek R1 671BThe full DeepSeek R1, for the biggest rigsneeds 406 GBToo big
deepseek-r1:671b
Every model the Apple M3 16 GB can run
4 of 12 fit. Best pick: Qwen3 4B.
- Qwen3 4BQuick everyday chat and summaries~22 tok/sUsable
qwen3:4b - Qwen3 8BA solid all-rounder for chat and writing~11 tok/sUsable
qwen3:8b - Gemma 3 12BReads images as well as text~7 tok/sSlow
gemma3:12b - Qwen3 14BNoticeably smarter at reasoning and code~6 tok/sSlow
qwen3:14b - gpt-oss 20BOpenAI's open model. Fast, strong reasoningneeds 16 GBToo big
gpt-oss:20b - Gemma 3 27BGoogle's best open model for its sizeneeds 19 GBToo big
gemma3:27b - Qwen3 30B-A3B30B smarts at small-model speedneeds 21 GBToo big
qwen3:30b - Qwen3 32BStrong coding and long answersneeds 22 GBToo big
qwen3:32b - Llama 3.3 70BClose to the big cloud modelsneeds 45 GBToo big
llama3.3:70b - gpt-oss 120BOpenAI's largest open modelneeds 67 GBToo big
gpt-oss:120b - Qwen3 235B-A22BAlibaba's flagship open model, for multi-GPU rigsneeds 144 GBToo big
qwen3:235b - DeepSeek R1 671BThe full DeepSeek R1, for the biggest rigsneeds 406 GBToo big
deepseek-r1:671b
Every model the Apple M3 24 GB can run
5 of 12 fit. Best pick: Qwen3 4B.
- Qwen3 4BQuick everyday chat and summaries~22 tok/sUsable
qwen3:4b - Qwen3 8BA solid all-rounder for chat and writing~11 tok/sUsable
qwen3:8b - Gemma 3 12BReads images as well as text~7 tok/sSlow
gemma3:12b - Qwen3 14BNoticeably smarter at reasoning and code~6 tok/sSlow
qwen3:14b - gpt-oss 20BOpenAI's open model. Fast, strong reasoning~18 tok/sUsable
gpt-oss:20b - Gemma 3 27BGoogle's best open model for its sizeneeds 19 GBToo big
gemma3:27b - Qwen3 30B-A3B30B smarts at small-model speedneeds 21 GBToo big
qwen3:30b - Qwen3 32BStrong coding and long answersneeds 22 GBToo big
qwen3:32b - Llama 3.3 70BClose to the big cloud modelsneeds 45 GBToo big
llama3.3:70b - gpt-oss 120BOpenAI's largest open modelneeds 67 GBToo big
gpt-oss:120b - Qwen3 235B-A22BAlibaba's flagship open model, for multi-GPU rigsneeds 144 GBToo big
qwen3:235b - DeepSeek R1 671BThe full DeepSeek R1, for the biggest rigsneeds 406 GBToo big
deepseek-r1:671b
Can the Apple M3 run it?
Can the Apple M3 run Qwen3 14B?
Yes, with 16 GB or more. Qwen3 14B needs about 11 GB and the Apple M3 with 16 GB of unified memory leaves about 11 GB for models. Expect about 6 tok/s, slow, but it works. The 8 GB version doesn't have room.
Can the Apple M3 run gpt-oss 20B?
Yes, with 24 GB or more. gpt-oss 20B needs about 16 GB and the Apple M3 with 24 GB of unified memory leaves about 16 GB for models. Expect about 18 tok/s, about reading speed. The 8 and 16 GB versions don't have room.
Can the Apple M3 run Gemma 3 27B?
No. Gemma 3 27B needs about 19 GB and the Apple M3 has at most 24 GB of unified memory, of which about 16 GB is usable for models. The most capable model it runs well is Qwen3 4B, at about 22 tok/s.
Real runs on the Apple M3
Loading measured runs…
Generating a token reads the model's active weights from memory once, so speed is roughly memory bandwidth (100 GB/s) divided by the active weight size, with efficiency calibrated on our own rack. Sizes are Ollama's default 4-bit downloads plus about 1.2 GB for context. macOS lets the GPU use about two thirds to three quarters of unified memory, so that is what we count. Real results are usually within about a third. Measured runs are on the leaderboard.
The Used workstation + 2× Tesla P100 runs Qwen3 32B at about 20 tok/s, from $750 used.