Best local AI models for NVIDIA NVS 810

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 NVS 810 is a specialized multi display graphics card equipped with 2 GB of DDR3 memory. In the context of local artificial intelligence, this VRAM capacity acts as a strict physical boundary. To run an AI model entirely on this hardware, the entire weight file and its operational memory must fit inside this 2 GB limit. Because DDR3 memory has lower bandwidth than modern GDDR memory, keeping the model small is critical for maintaining usable processing speeds.

To make models fit into this 2 GB memory space, we use quantized versions. The quantization column shows the optimal format for each model. For example, the Allegro 2.8B model fits by using the Q4_K_M quantization which reduces the model size so it uses exactly 2 GB of VRAM. Smaller models like SmolLM2 1.7B can use the higher quality Q6_K quantization while using 1.7 GB of VRAM. TinyLlama 1.1B runs at the highest Q8_0 quantization level using only 1.4 GB of VRAM.

When a model exceeds the 2 GB VRAM limit, you must use CPU offloading. This process splits the model weights between your graphics card and your system RAM. If you have a system with 32 GB of system RAM, you can run larger models like SmolLM3 3B or Kandinsky 3.1. These 3B models need 2.2 GB of VRAM at Q4_K_M quantization and require an additional 4.2 GB of system RAM to function.

Offloading comes with a significant performance cost. Moving data between the DDR3 VRAM of the NVIDIA NVS 810 and the system RAM over the PCIe bus creates a major bottleneck. While offloading allows you to run larger models like the 3.5B parameter SDXL Turbo or SDXL Lightning, the generation speed will be much slower than running a smaller model fully inside the graphics card memory.

You must also consider the memory required for context. Running text models with a standard 4k context window increases the memory footprint beyond the base model size. For a card with 2 GB of VRAM, this extra context memory can easily cause an out of memory error. When running models near the memory limit, you may need to reduce the context window to keep the system stable.