Best local AI models for NVIDIA GTX 650 Ti Boost

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 GTX 650 Ti Boost is an older graphics card equipped with 2 GB of GDDR5 memory. This physical memory limit dictates which artificial intelligence models you can run locally. To fit models onto this hardware, you must look at the memory size of the model and the quantization level. The memory size represents the physical footprint of the model parameters when loaded into your hardware.

The quant column shows the specific compression level used to shrink these models. Quantization reduces the precision of the model weights to save space. For example, the 2.8B Allegro model and the 2.7B Open-Sora Plan model both fit into your 2 GB of video memory when compressed to the Q4_K_M quantization level. Similarly, the 2.6B LFM2 and 2.6B Playground v2.5 models use 1.9 GB of video memory at this same Q4_K_M level.

As model sizes decrease, you can use higher quality quantization levels. The 2.3B SeamlessM4T v2 and 2.2B Kimi K3 DSpark models run at the Q5_K_M level while using exactly 2 GB of video memory. The 2B SmolVLM, 2B Stable Diffusion 3 Medium, and 2B Pyramid Flow models can run at the Q6_K level. Smaller models like the 1.6B StableLM 2 and 1.6B Sana use the Q8_0 level, which preserves more original model quality.

Running models on a card with 2 GB of memory comes with a strict context caveat. Standard calculations often assume a basic context window. If you increase your context window to 4k tokens, the memory required for processing text history will grow quickly. This extra memory demand can easily exceed your remaining video memory and cause out of memory errors.

When a model is too large for your video memory, you must use CPU offload. This technique splits the model weights between your graphics card and your system RAM. We assume your computer has 32 GB of system RAM for these scenarios. Offloading allows you to run larger models, but it costs a lot of processing speed because transferring data between system RAM and your graphics card is slow.

With CPU offload, you can run the 3B SmolLM3 or 3B Kandinsky 3.1 models. These models need 2.2 GB of video memory at the Q4_K_M level and require 4.2 GB of system RAM. You can also run the 3.5B SDXL Turbo or 3.5B SDXL Lightning models. These models require 2.6 GB of video memory at the Q4_K_M level and 4.6 GB of system RAM to function.