RunsLocalGPU checker Local AI toolsv2.2 · data Oct 17, 2026

What AI models can the M5 Max 36GB run?

With 36 GB unified memory, the M5 Max 36GB runs 17 of 25 popular open models fully in fast memory at Q4_K_M or better (8K context). The largest that fit are Hermes 4.3 36B (Q4_K_M), Qwen3.6 35B-A3B (Q5_K_M), Gemma 4 31B (Q5_K_M). Fastest: Qwen3.6 35B-A3B at about 129 tokens/sec and Gemma 4 E2B at about 113 tokens/sec.

Check with your own RAM and context length

Model	Result	Best quant	Memory needed	Est. speed
Qwen3.6 35B-A3B MoE · coding	Runs on GPU	Q5_K_M	25.9 GB	129 tok/s
Gemma 4 E2B small	Runs on GPU	Q8_0	6.1 GB	113 tok/s
Nemotron 3.5 Lightning 30B-A3B MoE · agentic	Runs on GPU	Q6_K	25.4 GB	112 tok/s
Mellum2.1 12B-A2.5B Thinking MoE · coding	Runs on GPU	Q8_0	13.3 GB	104 tok/s
Gemma 4 26B A4B MoE · vision	Runs on GPU	Q6_K	21.8 GB	89 tok/s
gpt-oss-20b MoE · reasoning	Runs on GPU	Q8_0	23.2 GB	72 tok/s
Ministral 3 3B small	Runs on GPU	Q8_0	4.8 GB	68 tok/s
Gemma 4 E4B small · audio	Runs on GPU	Q8_0	9.3 GB	58 tok/s
Qwen3.5 9B general	Runs on GPU	Q8_0	11.3 GB	27 tok/s
Gemma 4 12B general · audio	Runs on GPU	Q8_0	13.8 GB	22 tok/s
Phi-4 14B reasoning	Runs on GPU	Q8_0	17.7 GB	18 tok/s
Mistral Small 3.1 24B general	Runs on GPU	Q6_K	21.5 GB	14 tok/s
Gemma 4 31B general · vision	Runs on GPU	Q5_K_M	23.0 GB	13 tok/s
Hermes 4.3 36B general	Runs on GPU	Q4_K_M	23.9 GB	13 tok/s
Qwen3.6 27B coding	Runs on GPU	Q6_K	23.2 GB	12 tok/s
Qwen3.8 27B coding	Runs on GPU	Q6_K	23.9 GB	12 tok/s
Granite 4.2 30B general	Runs on GPU	Q6_K	24.8 GB	11 tok/s
Llama 3.3 70B general	Too big	Q4_K_M	45.9 GB	—
Llama 4 Scout MoE · vision	Too big	Q4_K_M	68.1 GB	—
gpt-oss-120b MoE · reasoning	Too big	Q4_K_M	72.0 GB	—
Qwen3.5 122B-A10B MoE	Too big	Q4_K_M	75.3 GB	—
Qwen3.8-Flash-Next MoE	Too big	Q4_K_M	107.8 GB	—
GLM-5.3-Flash MoE · coding	Too big	Q4_K_M	196.0 GB	—
DeepSeek V4.1-Flash MoE · coding	Too big	Q4_K_M	335.8 GB	—
Muse Spark Open 816B-A42B MoE · vision	Too big	Q4_K_M	497.3 GB	—

Specs used: 36 GB unified memory, 460 GB/s memory bandwidth. Speeds are estimates for generating text with llama.cpp-style runtimes; real numbers usually land within about 20%.

Apps that support the M5 Max 36GB

Ollama LM Studio Jan GPT4All AnythingLLM llama.cpp MLX LM KoboldCpp Text Generation WebUI LocalAI