Which model should you run?
Ordered from smallest to most capable. Each page says what the model is good for, how much memory it needs, and every GPU and Mac that runs it.
- Qwen3 4BAlibaba (Qwen) · Apache 2.0 · about 4 GBQuick everyday chat and summaries.
- Qwen3 8BAlibaba (Qwen) · Apache 2.0 · about 7 GBA solid all-rounder for chat and writing.
- Gemma 3 12BGoogle DeepMind · Gemma Terms of Use · about 10 GBReads images as well as text.
- Qwen3 14BAlibaba (Qwen) · Apache 2.0 · about 11 GBNoticeably smarter at reasoning and code.
- gpt-oss 20BOpenAI · Apache 2.0 · about 16 GBOpenAI's open model. Fast, strong reasoning.
- Gemma 3 27BGoogle DeepMind · Gemma Terms of Use · about 19 GBGoogle's best open model for its size.
- Qwen3 30B-A3BAlibaba (Qwen) · Apache 2.0 · about 21 GB30B smarts at small-model speed.
- Qwen3 32BAlibaba (Qwen) · Apache 2.0 · about 22 GBStrong coding and long answers.
- Llama 3.3 70BMeta · Llama 3.3 Community License · about 45 GBClose to the big cloud models.
- gpt-oss 120BOpenAI · Apache 2.0 · about 67 GBOpenAI's largest open model.
- Qwen3 235B-A22BAlibaba (Qwen) · Apache 2.0 · about 144 GBAlibaba's flagship open model, for multi-GPU rigs.
- DeepSeek R1 671BDeepSeek · MIT · about 406 GBThe full DeepSeek R1, for the biggest rigs.