Skip to content
Run LLM Locally

Mac · Apple Silicon

M2 with 16 GB: which local LLMs run?

With 16 GB of unified memory, the M2 gives the GPU about 10.7 GB by default (100 GB/s): enough for models up to about Mistral Nemo 12B Instruct.

  • Unified memory16 GB
  • GPU usable~10.7 GB
  • Bandwidth100 GB/s
  • GPU cores10
ShareFacebookXWhatsAppLinkedInEmail

Unified memory is not VRAM: macOS reserves part of it. The limit can be raised with sudo sysctl iogpu.wired_limit_mb=…, at the cost of memory for apps. There is no offload to system RAM — it is the same memory.

CPU cores
8
GPU cores
10
Bandwidth
100 GB/s
Memory configurations
8 GB · 16 GB · 24 GB
Sold in
MacBook Air, MacBook Pro 13, Mac mini
Released
June 2022

Official specData updated 24 days agoSource: Apple tech specs

What the M2 with 16 GB runs

M2 · 16 GB · 8K
Gemma 3 4B ITQ8_0 · 8K~5.2 GB · 15–22 tok/s estimateExcellent
Qwen3 4BQ8_0 · 8K~6 GB · 13–19 tok/s estimateExcellent
Phi-4 Mini Instruct 3.8BQ8_0 · 8K~5.8 GB · 14–20 tok/s estimateExcellent
Llama 3.2 3B InstructQ8_0 · 8K~4.9 GB · 17–24 tok/s estimateExcellent
Qwen3 1.7BQ8_0 · 8K~3.5 GB · 27–40 tok/s estimateExcellent
Llama 3.2 1B InstructQ8_0 · 8K~2.3 GB · 43–63 tok/s estimateExcellent
Qwen3 0.6BQ8_0 · 8K~2.3 GB · 51–83 tok/s estimateExcellent
Qwen3 14BIQ4_XS · 8K~9.9 GB · 7.5–11 tok/s estimateCompatible
DeepSeek R1 Distill Qwen 14BIQ4_XS · 8K~10.2 GB · 7–11 tok/s estimateCompatible
Qwen2.5 14B InstructIQ4_XS · 8K~10.2 GB · 7–11 tok/s estimateCompatible
Phi-4 14BIQ4_XS · 8K~10 GB · 7.5–11 tok/s estimateCompatible
Mistral Nemo 12B InstructQ4_K_M · 8K~9.2 GB · 8–12 tok/s estimateCompatible
Gemma 3 12B ITQ5_K_M · 8K~10 GB · 7–11 tok/s estimateCompatible
Gemma 2 9B ITQ5_K_M · 8K~9.4 GB · 8.5–12 tok/s estimateCompatible
Qwen3 8BQ6_K · 8K~8.4 GB · 9–13 tok/s estimateCompatible
Llama 3.1 8B InstructQ8_0 · 8K~10 GB · 7–10 tok/s estimateCompatible
DeepSeek R1 Distill Llama 8BQ8_0 · 8K~10 GB · 7–10 tok/s estimateCompatible
Qwen2.5 7B InstructQ8_0 · 8K~9 GB · 8–11 tok/s estimateCompatible
Mistral 7B Instruct v0.3Q8_0 · 8K~9 GB · 8–11 tok/s estimateCompatible
DeepSeek R1 (671B)Q8_0 · 8K~699 GB · No dataIncompatible
Qwen3 235B-A22BQ8_0 · 8K~247 GB · No dataIncompatible
gpt-oss-120bMXFP4 · 8K~63 GB · No dataIncompatible
Llama 4 Scout 17B-16E InstructQ8_0 · 8K~113 GB · No dataIncompatible
Qwen2.5 72B InstructQ8_0 · 8K~79.1 GB · No dataIncompatible
Llama 3.3 70B InstructQ8_0 · 8K~76.6 GB · No dataIncompatible
DeepSeek R1 Distill Llama 70BQ8_0 · 8K~76.6 GB · No dataIncompatible
Mixtral 8x7B InstructQ8_0 · 8K~50 GB · No dataIncompatible
Qwen3 32BQ8_0 · 8K~36.8 GB · No dataIncompatible
DeepSeek R1 Distill Qwen 32BQ8_0 · 8K~36.8 GB · No dataIncompatible
Qwen2.5 32B InstructQ8_0 · 8K~36.8 GB · No dataIncompatible
Qwen2.5 Coder 32B InstructQ8_0 · 8K~36.8 GB · No dataIncompatible
QwQ 32BQ8_0 · 8K~36.8 GB · No dataIncompatible
Qwen3 30B-A3BQ8_0 · 8K~33.2 GB · No dataIncompatible
Qwen3 Coder 30B-A3B InstructQ8_0 · 8K~33.2 GB · No dataIncompatible
Gemma 3 27B ITQ8_0 · 8K~30 GB · No dataIncompatible
Gemma 2 27B ITQ8_0 · 8K~31.4 GB · No dataIncompatible
Mistral Small 3.2 24B InstructQ8_0 · 8K~26.4 GB · No dataIncompatible
gpt-oss-20bMXFP4 · 8K~12.8 GB · No dataIncompatible

For each model, the highest-precision quantization that runs entirely in fast memory; otherwise the one with the best result. Speeds are ranges; "estimate" unless a benchmark exists for this exact hardware.

EstimateData updated 24 days agoSource: Run LLM Locally engine