Best local AI models for NVIDIA GT 635M

The largest models that fit fully

The 30 largest of the 56 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 NVIDIA GT 635M is an entry level mobile graphics card equipped with 2 GB of DDR3 video memory. This dedicated memory size determines the maximum size of the AI models you can run entirely on the hardware. Because the card has 2 GB of VRAM, any model running fully on the GPU must fit within this strict limit. This hardware is best suited for highly optimized small language models, compact image generators, and lightweight audio tools.

To make models fit into this limited memory, developers use quantization. The quant column shows the specific level of compression applied to the model weights. For example, a Q4_K_M quant uses approximately four bits per weight, which allows larger models like Allegro 2.8B or Open-Sora Plan 2.7B to fit into 2 GB of VRAM. Higher quants like Q6_K or Q8_0 offer better output quality but require more memory per parameter, limiting you to smaller models like Moondream 2 1.9B or TinyLlama 1.1B.

When a model exceeds the 2 GB VRAM limit, you must use CPU offload. This technique splits the model layers between your GPU memory and your system RAM. If you have 32 GB of system RAM, you can run larger models like SmolLM3 3B or Kandinsky 3.1. These models need 2.2 GB of VRAM at Q4_K_M and require an additional 4.2 GB of system RAM. Offloading allows you to run these larger architectures, but it comes at a cost. Transferring data between system RAM and your GPU over the system bus significantly reduces processing speeds.

For image generation, CPU offload allows you to run demanding models such as Stable Diffusion XL. This 3.417B parameter model needs 4.1 GB at FP8 or optimized settings, which requires 6.1 GB of system RAM to supplement your VRAM. Similarly, SDXL Turbo and SDXL Lightning require 2.6 GB of VRAM at Q4_K_M and 4.6 GB of system RAM. While these setups function, generation times will be much slower than fully on card execution.

You must also consider the context window when running text models. The memory figures listed are for the model weights alone. Running a model with a standard 4k context window requires additional VRAM to store the active conversation history. On a 2 GB card, loading a model that uses almost the entire memory limit leaves no room for this context. To avoid running out of memory during generation, you should select smaller models like Qwen3 1.7B or SmolLM2 1.7B to reserve space for your active text context.