Best local AI models for Intel Arc A770M

The largest models that fit fully

The 30 largest of the 155 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.

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.

How to read this

The Intel Arc A770M is a mobile graphics processor equipped with 16 GB of GDDR6 dedicated memory. This memory capacity determines which artificial intelligence models you can run entirely on your hardware. When a model fits completely within the graphics memory, it processes tokens at maximum speed. If a model exceeds this limit, some data must spill over into your system memory, which slows down performance.

To fit larger models into the available space, developers use quantization. The quantization column shows the compression format used to reduce the size of the model weights. For example, the 21B gpt-oss-20b and Reka Flash 3 models fit into 15.4 GB of memory when compressed to the Q4_K_M format. Similarly, the 20B Qwen-Image and Qwen-Image-Edit models fit into 14.6 GB of memory using the same Q4_K_M quantization.

As model sizes decrease, you can use higher quality quantization formats. The 17B HunyuanImage 2.1 / 3.0 fits into 14.5 GB of memory using the Q5_K_M format. You can run the 16B DeepSeek-Coder-V2 16B / 236B and Kimi-VL A3B models at Q6_K quantization, which uses 15.7 GB of memory. The 14B models like Phi-4, Phi-4-reasoning / -plus, Wan 2.2 T2I, and SkyReels V2 also run at Q6_K quantization, consuming 13.8 GB of memory.

For even better precision, 12B models can run at the Q8_0 quantization level. The Gemma 3 12B, Gemma 4 12B, Mistral NeMo 12B, and Pixtral 12B models all use 15.3 GB of memory at Q8_0. Image generation models like FLUX.1 dev can run on an optimized FP8 quantization that uses 14.4 GB of memory. These configurations maximize the output quality while remaining within the hardware limits.

If you want to run larger models, you must use CPU offloading. This process shares the workload between your graphics card and your system RAM. Assuming you have 32 GB of system RAM, you can run the 22B Solar Pro or Codestral 22B models at Q4_K_M, which require 16.1 GB of memory and 18.1 GB of system RAM. The 27B Gemma 3 27B model requires 19.8 GB of memory and 21.8 GB of system RAM at Q4_K_M. Offloading allows these models to run, but it reduces processing speed.

When planning your setup, remember that context length affects memory usage. The memory figures listed here are calculated using a standard 4k context window. If you increase the context window to process longer documents or conversations, the system will require more memory. This extra demand might force you to use a smaller model or a lower quantization level to prevent system slowdowns.