Open models
Open-weight models, with what each needs to run on a Mac.
Runs on your MacGoogle DeepMindGemma 4 31BGoogle's largest dense Gemma 4 model: text and image input, a 256K-token context and over 140 languages.
ViewRuns on your MacOpenAIgpt-oss-20bOpenAI's smaller open-weight reasoning model: 21B parameters with 3.6B active, with adjustable reasoning effort.
View
ViewRuns on your MacOpenAIgpt-oss-20bOpenAI's smaller open-weight reasoning model: 21B parameters with 3.6B active, with adjustable reasoning effort.ViewAll Models
Best MLX Score first.
16 of 16
1Gemma 4 31BGoogle's largest dense Gemma 4 model: text and image input, a 256K-token context and over 140 languages. 5.0View2gpt-oss-20bOpenAI's smaller open-weight reasoning model: 21B parameters with 3.6B active, with adjustable reasoning effort. 4.7View3
Llama 3.1 8B InstructMeta's 8B multilingual chat model with a 128K-token context; small enough for an 8 GB Mac at 4 bits. 4.5View4
Qwen3.5 9BA 9B model from Alibaba that reads text, images and video, covers 201 languages and fits on a 16 GB Mac. 4.2View5
Qwen3.6 35B A3BAlibaba's 35B mixture-of-experts model with 3B active parameters, for chat and agentic coding; reads text, images and video. 3.9View6
DeepSeek V4 Flash 0731DeepSeek's 284B mixture-of-experts model with 13B active and a 1M-token context, for Macs with 256 GB of memory. 3.7View7
gpt-oss-120bOpenAI's larger open-weight reasoning model: 117B parameters with 5.1B active, for Macs with 96 GB of memory or more. 3.4View8
Gemma 4 E4BA small Gemma 4 model that takes text, images and audio, with a 128K-token context; fits an 8 GB Mac at 4 bits. 3.1View9
Qwen3.6 27BAlibaba's 27B dense model for chat and agentic coding, reading text, images and video, with a 262K-token context. 2.9View10
Nemotron 3 Nano 30B A3BNVIDIA's hybrid Mamba-Transformer mixture-of-experts model: 30B parameters, 3.5B active, reasoning you can switch off. 2.6View11
GLM-4.7 FlashZ.ai's 30B-A3B mixture-of-experts model for lightweight deployment, with a preserved thinking mode for agent tasks. 2.3View12
Qwen3 Coder 30B A3B InstructAlibaba's coding model: 30.5B parameters with 3.3B active, made for agentic coding, with a 256K-token context. 2.1View13
Llama 3.3 70B InstructMeta's 70B multilingual text model with a 128K-token context, for Macs with 64 GB of memory or more. 1.8View14
Phi-4 mini instructMicrosoft's 3.8B model focused on reasoning, maths and logic, with a 128K-token context; about 2 GB at 4 bits. 1.5View15
Devstral Small 2 (24B)Mistral's 24B agentic coding model for exploring codebases and editing many files; Mistral says it runs on a 32 GB Mac. 1.3View16
Mistral Small 3.2 (24B)Mistral's 24B instruct model with image input and function calling; fits a 24 GB Mac at 4 bits. 1.0View
Gemma 4 31BGoogle's largest dense Gemma 4 model: text and image input, a 256K-token context and over 140 languages. 5.0View2gpt-oss-20bOpenAI's smaller open-weight reasoning model: 21B parameters with 3.6B active, with adjustable reasoning effort. 4.7View3Llama 3.1 8B InstructMeta's 8B multilingual chat model with a 128K-token context; small enough for an 8 GB Mac at 4 bits. 4.5View4Qwen3.5 9BA 9B model from Alibaba that reads text, images and video, covers 201 languages and fits on a 16 GB Mac. 4.2View5Qwen3.6 35B A3BAlibaba's 35B mixture-of-experts model with 3B active parameters, for chat and agentic coding; reads text, images and video. 3.9View6DeepSeek V4 Flash 0731DeepSeek's 284B mixture-of-experts model with 13B active and a 1M-token context, for Macs with 256 GB of memory. 3.7View7gpt-oss-120bOpenAI's larger open-weight reasoning model: 117B parameters with 5.1B active, for Macs with 96 GB of memory or more. 3.4View8Gemma 4 E4BA small Gemma 4 model that takes text, images and audio, with a 128K-token context; fits an 8 GB Mac at 4 bits. 3.1View9Qwen3.6 27BAlibaba's 27B dense model for chat and agentic coding, reading text, images and video, with a 262K-token context. 2.9View10Nemotron 3 Nano 30B A3BNVIDIA's hybrid Mamba-Transformer mixture-of-experts model: 30B parameters, 3.5B active, reasoning you can switch off. 2.6View11GLM-4.7 FlashZ.ai's 30B-A3B mixture-of-experts model for lightweight deployment, with a preserved thinking mode for agent tasks. 2.3View12Qwen3 Coder 30B A3B InstructAlibaba's coding model: 30.5B parameters with 3.3B active, made for agentic coding, with a 256K-token context. 2.1View13Llama 3.3 70B InstructMeta's 70B multilingual text model with a 128K-token context, for Macs with 64 GB of memory or more. 1.8View14Phi-4 mini instructMicrosoft's 3.8B model focused on reasoning, maths and logic, with a 128K-token context; about 2 GB at 4 bits. 1.5View15Devstral Small 2 (24B)Mistral's 24B agentic coding model for exploring codebases and editing many files; Mistral says it runs on a 32 GB Mac. 1.3View16Mistral Small 3.2 (24B)Mistral's 24B instruct model with image input and function calling; fits a 24 GB Mac at 4 bits. 1.0View