Best local AI models for Apple M2 Pro
11.2 GB usable of 16 GB unified memory. At a 4k context, 144 of the 233 models in our catalog with verified parameter counts fit fully, up to Apriel-1.5-15B-Thinker at 15B parameters. Computed for the 16 GB configuration; a larger memory configuration fits more.
Check your own machine against every model →The largest models that fit fully
The 30 largest of the 144 models that fit; every smaller model in the catalog fits too. Best quant means the highest quality compression whose weights and 4k context both sit inside the memory.
| Model | Parameters | Best quant that fits | Memory used at 4k |
|---|---|---|---|
| Apriel-1.5-15B-Thinker | 15B | Q4_K_M | 11 GB |
| StarCoder2 3B / 7B / 15B | 15B | Q4_K_M | 11 GB |
| Phi-3 Medium | 14B | Q4_K_M | 10.2 GB |
| Phi-4 | 14B | Q4_K_M | 10.2 GB |
| Phi-4-reasoning / -plus | 14B | Q4_K_M | 10.2 GB |
| Wan 2.2 T2I | 14B | Q4_K_M | 10.2 GB |
| Wan 2.1 (1.3B / 14B) | 14B | Q4_K_M | 10.2 GB |
| SkyReels V2 | 14B | Q4_K_M | 10.2 GB |
| Vicuna 13B | 13B | Q5_K_M | 11.1 GB |
| HunyuanVideo | 13B | Q5_K_M | 11.1 GB |
| HunyuanVideo-Avatar | 13B | Q5_K_M | 11.1 GB |
| LTX-Video / LTX-2 | 13B | Q5_K_M | 11.1 GB |
| FramePack | 13B | Q5_K_M | 11.1 GB |
| Gemma 3 12B | 12B | Q5_K_M | 10.2 GB |
| Gemma 4 12B | 12B | Q5_K_M | 10.2 GB |
| Mistral NeMo 12B | 12B | Q5_K_M | 10.2 GB |
| Pixtral 12B | 12B | Q5_K_M | 10.2 GB |
| FLUX.1 schnell | 12B | Q5_K_M | 10.2 GB |
| FLUX.1 Kontext dev | 12B | Q5_K_M | 10.2 GB |
| FLUX.1 Krea dev | 12B | Q5_K_M | 10.2 GB |
| Open-Sora 2.0 | 11B | Q6_K | 10.8 GB |
| Mochi 1 | 10B | Q6_K | 9.8 GB |
| Gemma 2 9B | 9B | Q6_K | 10.3 GB |
| Nemotron Nano 4B / 9B | 9B | Q6_K | 8.9 GB |
| GLM-4 9B / GLM-4.5-Air | 9B | Q6_K | 8.9 GB |
| Yi-Coder 1.5B / 9B | 9B | Q6_K | 8.9 GB |
| GLM-4-9B-Chat / CodeGeeX4 | 9B | Q6_K | 8.9 GB |
| GLM-4V-9B / GLM-4.1V-Thinking | 9B | Q6_K | 8.9 GB |
| Chroma | 8.9B | Q6_K | 8.8 GB |
| Llama 3.1 8B | 8B | Q8_0 | 10.7 GB |
Close, but only with CPU offload
These need more than the card holds at their smallest practical quant, so part of the model runs from system memory (figures assume 32 GB of it). They work, several times slower.
| Model | Parameters | Memory at FP8 / optimized | System RAM at 4k |
|---|---|---|---|
| FLUX.1 dev | 12B | 14.4 GB needed | 16.4 GB |
| Qwen2.5 14B | 14.7B | 11.6 GB needed | 13.6 GB |
| DeepSeek-Coder-V2 16B / 236B | 16B | 11.7 GB needed | 13.7 GB |
| Kimi-VL A3B | 16B | 11.7 GB needed | 13.7 GB |
| Ling-Coder-Lite | 16.8B | 12.3 GB needed | 14.3 GB |
| HunyuanImage 2.1 / 3.0 | 17B | 12.4 GB needed | 14.4 GB |
| CogVLM2 | 19B | 13.9 GB needed | 15.9 GB |
| Qwen-Image | 20B | 14.6 GB needed | 16.6 GB |
| Qwen-Image-Edit | 20B | 14.6 GB needed | 16.6 GB |
| gpt-oss-20b | 21B | 15.4 GB needed | 17.4 GB |
How to read this
The Apple M2 Pro chip with a 16 GB unified memory pool provides 11.2 GB of usable memory for local AI models. This usable share is the maximum space available to load and run models entirely on the graphics hardware. If a model exceeds this limit it will not fit in the graphics memory. This page helps you choose the best models and quantization levels to run within this hardware boundary.
Quantization is a method that compresses model weights to save space. The quant column shows the best compression level for each model. For example the Apriel-1.5-15B-Thinker and StarCoder2 15B models use the Q4_K_M quant which requires 11 GB of memory. Models like Phi-3 Medium Phi-4 Phi-4-reasoning Phi-4-plus Wan 2.2 T2I Wan 2.1 14B and SkyReels V2 also use the Q4_K_M quant and require 10.2 GB of memory.
Smaller models can run at higher precision levels with less compression. The Vicuna 13B HunyuanVideo HunyuanVideo-Avatar LTX-Video LTX-2 and FramePack models fit using the Q5_K_M quant which uses 11.1 GB. Gemma 3 12B Gemma 4 12B Mistral NeMo 12B Pixtral 12B FLUX.1 schnell FLUX.1 Kontext dev and FLUX.1 Krea dev use the Q5_K_M quant and require 10.2 GB. Open-Sora 2.0 uses the Q6_K quant and requires 10.8 GB.
Highly compressed options allow excellent performance on smaller architectures. Mochi 1 uses the Q6_K quant and requires 9.8 GB. Gemma 2 9B uses Q6_K and requires 10.3 GB. Nemotron Nano 9B GLM-4 9B GLM-4.5-Air Yi-Coder 9B GLM-4-9B-Chat CodeGeeX4 GLM-4V-9B and GLM-4.1V-Thinking use Q6_K and require 8.9 GB. Chroma uses Q6_K and requires 8.8 GB. Llama 3.1 8B runs at the high quality Q8_0 quant using 10.7 GB.
When a model is too large for the 11.2 GB unified memory limit you must offload parts of it to the system CPU. This CPU offload requires a 32 GB system RAM configuration. Offloading allows you to run larger models but it reduces processing speed because the system RAM and CPU are slower than the unified graphics memory.
Examples of CPU offload models include FLUX.1 dev which needs 14.4 GB at FP8 and uses 16.4 GB of system RAM. Qwen2.5 14B needs 11.6 GB at Q4_K_M and uses 13.6 GB of system RAM. DeepSeek-Coder-V2 16B and Kimi-VL A3B need 11.7 GB at Q4_K_M and use 13.7 GB of system RAM. Ling-Coder-Lite needs 12.3 GB at Q4_K_M and uses 14.3 GB of system RAM. HunyuanImage 2.1 and HunyuanImage 3.0 need 12.4 GB at Q4_K_M and use 14.4 GB of system RAM.
Other offload options include CogVLM2 which needs 13.9 GB at Q4_K_M and uses 15.9 GB of system RAM. Qwen-Image and Qwen-Image-Edit need 14.6 GB at Q4_K_M and use 16.6 GB of system RAM. The gpt-oss-20b model needs 15.4 GB at Q4_K_M and uses 17.4 GB of system RAM. Note that all memory calculations assume a standard 4k context window. Expanding the context window beyond 4k tokens will require more memory and may force you to use smaller models.