← Every GPUApple · 96 or 256 or 512 GB unified memory · 819 GB/s

What can the Apple M3 Ultra run?

With 96 GB, 10 of 12 popular open models fit. The best pick is gpt-oss 120B at about 99 tok/s, faster than you can read.

ollama run gpt-oss:120b

First time? Install Ollama, then paste that into a terminal. Is this your machine? Check it on the homepage.

96 GB unified memory · about 72 GB for models

Every model the Apple M3 Ultra 96 GB can run

10 of 12 fit. Best pick: gpt-oss 120B.

  • Qwen3 4BQuick everyday chat and summaries
    ~180 tok/s
    Runs greatqwen3:4b
  • Qwen3 8BA solid all-rounder for chat and writing
    ~87 tok/s
    Runs greatqwen3:8b
  • Gemma 3 12BReads images as well as text
    ~56 tok/s
    Runs greatgemma3:12b
  • Qwen3 14BNoticeably smarter at reasoning and code
    ~48 tok/s
    Runs greatqwen3:14b
  • gpt-oss 20BOpenAI's open model. Fast, strong reasoning
    ~140 tok/s
    Runs greatgpt-oss:20b
  • Gemma 3 27BGoogle's best open model for its size
    ~26 tok/s
    Usablegemma3:27b
  • Qwen3 30B-A3B30B smarts at small-model speed
    ~140 tok/s
    Runs greatqwen3:30b
  • Qwen3 32BStrong coding and long answers
    ~23 tok/s
    Usableqwen3:32b
  • Llama 3.3 70BClose to the big cloud models
    ~10 tok/s
    Usablellama3.3:70b
  • gpt-oss 120BOpenAI's largest open model
    ~99 tok/s
    Runs greatgpt-oss:120b
  • Qwen3 235B-A22BAlibaba's flagship open model, for multi-GPU rigs
    needs 144 GB
    Too bigqwen3:235b
  • DeepSeek R1 671BThe full DeepSeek R1, for the biggest rigs
    needs 406 GB
    Too bigdeepseek-r1:671b
256 GB unified memory · about 192 GB for models

Every model the Apple M3 Ultra 256 GB can run

11 of 12 fit. Best pick: Qwen3 235B-A22B.

  • Qwen3 4BQuick everyday chat and summaries
    ~180 tok/s
    Runs greatqwen3:4b
  • Qwen3 8BA solid all-rounder for chat and writing
    ~87 tok/s
    Runs greatqwen3:8b
  • Gemma 3 12BReads images as well as text
    ~56 tok/s
    Runs greatgemma3:12b
  • Qwen3 14BNoticeably smarter at reasoning and code
    ~48 tok/s
    Runs greatqwen3:14b
  • gpt-oss 20BOpenAI's open model. Fast, strong reasoning
    ~140 tok/s
    Runs greatgpt-oss:20b
  • Gemma 3 27BGoogle's best open model for its size
    ~26 tok/s
    Usablegemma3:27b
  • Qwen3 30B-A3B30B smarts at small-model speed
    ~140 tok/s
    Runs greatqwen3:30b
  • Qwen3 32BStrong coding and long answers
    ~23 tok/s
    Usableqwen3:32b
  • Llama 3.3 70BClose to the big cloud models
    ~10 tok/s
    Usablellama3.3:70b
  • gpt-oss 120BOpenAI's largest open model
    ~99 tok/s
    Runs greatgpt-oss:120b
  • Qwen3 235B-A22BAlibaba's flagship open model, for multi-GPU rigs
    ~22 tok/s
    Usableqwen3:235b
  • DeepSeek R1 671BThe full DeepSeek R1, for the biggest rigs
    needs 406 GB
    Too bigdeepseek-r1:671b
512 GB unified memory · about 384 GB for models

Every model the Apple M3 Ultra 512 GB can run

11 of 12 fit. Best pick: Qwen3 235B-A22B.

  • Qwen3 4BQuick everyday chat and summaries
    ~180 tok/s
    Runs greatqwen3:4b
  • Qwen3 8BA solid all-rounder for chat and writing
    ~87 tok/s
    Runs greatqwen3:8b
  • Gemma 3 12BReads images as well as text
    ~56 tok/s
    Runs greatgemma3:12b
  • Qwen3 14BNoticeably smarter at reasoning and code
    ~48 tok/s
    Runs greatqwen3:14b
  • gpt-oss 20BOpenAI's open model. Fast, strong reasoning
    ~140 tok/s
    Runs greatgpt-oss:20b
  • Gemma 3 27BGoogle's best open model for its size
    ~26 tok/s
    Usablegemma3:27b
  • Qwen3 30B-A3B30B smarts at small-model speed
    ~140 tok/s
    Runs greatqwen3:30b
  • Qwen3 32BStrong coding and long answers
    ~23 tok/s
    Usableqwen3:32b
  • Llama 3.3 70BClose to the big cloud models
    ~10 tok/s
    Usablellama3.3:70b
  • gpt-oss 120BOpenAI's largest open model
    ~99 tok/s
    Runs greatgpt-oss:120b
  • Qwen3 235B-A22BAlibaba's flagship open model, for multi-GPU rigs
    ~22 tok/s
    Usableqwen3:235b
  • DeepSeek R1 671BThe full DeepSeek R1, for the biggest rigs
    needs 406 GB
    Too bigdeepseek-r1:671b
Common questions

Can the Apple M3 Ultra run it?

Can the Apple M3 Ultra run gpt-oss 120B?

Yes. gpt-oss 120B needs about 67 GB and the Apple M3 Ultra with 96 GB of unified memory leaves about 72 GB for models. Expect about 99 tok/s, faster than you can read.

Can the Apple M3 Ultra run Qwen3 235B-A22B?

Yes, with 256 GB or more. Qwen3 235B-A22B needs about 144 GB and the Apple M3 Ultra with 256 GB of unified memory leaves about 192 GB for models. Expect about 22 tok/s, about reading speed. The 96 GB version doesn't have room.

Can the Apple M3 Ultra run DeepSeek R1 671B?

No. DeepSeek R1 671B needs about 406 GB and the Apple M3 Ultra has at most 512 GB of unified memory, of which about 384 GB is usable for models. The most capable model it runs well is Qwen3 235B-A22B, at about 22 tok/s.

Measured, not estimated

Real runs on the Apple M3 Ultra

Loading measured runs…

How we estimate

Generating a token reads the model's active weights from memory once, so speed is roughly memory bandwidth (819 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.

Want something bigger?

The Dell DSS 8440 + 10× Tesla P100 runs Qwen3 235B-A22B at about 19 tok/s, from $5,100 used.

See the builds

Compare

Similar cards

Every GPU we track

Head to head

The Apple M3 Ultra against its rivals

Every comparison