Best local AI models for NVIDIA NVS 310
0.5 GB DDR3. At a 4k context, 12 of the 233 models in our catalog with verified parameter counts fit fully, up to Qwen3 0.6B at 0.6B parameters.
Check your own machine against every model →The largest models that fit fully
The 12 largest of the 12 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 |
|---|---|---|---|
| Qwen3 0.6B | 0.6B | Q5_K_M | 0.5 GB |
| PixArt-α / PixArt-Σ | 0.6B | Q5_K_M | 0.5 GB |
| Parakeet TDT 0.6B v2 | 0.6B | Q5_K_M | 0.5 GB |
| XTTS v2 | 0.5B | Q6_K | 0.5 GB |
| Spark-TTS | 0.5B | Q6_K | 0.5 GB |
| CosyVoice 2 | 0.5B | Q6_K | 0.5 GB |
| VALL-E X (unofficial) | 0.4B | Q8_0 | 0.5 GB |
| ERNIE 4.5 open weights | 0.3B | Q8_0 | 0.4 GB |
| F5-TTS | 0.3B | Q8_0 | 0.4 GB |
| E2-TTS | 0.3B | Q8_0 | 0.4 GB |
| ChatTTS | 0.3B | Q8_0 | 0.4 GB |
| StyleTTS 2 | 0.15B | FP16 | 0.4 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 Q4_K_M | System RAM at 4k |
|---|---|---|---|
| Florence-2 base/large | 0.77B | 0.6 GB needed | 2.6 GB |
| Magenta RT | 0.8B | 0.6 GB needed | 2.6 GB |
| Stable Diffusion 2.1 | 0.9B | 0.7 GB needed | 2.7 GB |
| Bark | 0.9B | 0.7 GB needed | 2.7 GB |
| Tortoise TTS | 0.9B | 0.7 GB needed | 2.7 GB |
| Riffusion (SD-based) | 0.9B | 0.7 GB needed | 2.7 GB |
| Gemma 3 1B | 1B | 0.7 GB needed | 2.7 GB |
| Llama 3.2 1B / 3B | 1B | 0.7 GB needed | 2.7 GB |
| MMS (1100+ languages) | 1B | 0.7 GB needed | 2.7 GB |
| CSM-1B | 1B | 0.7 GB needed | 2.7 GB |
How to read this
The NVIDIA NVS 310 is an entry level graphics card equipped with 0.5 GB of DDR3 video memory. This very small VRAM capacity severely limits local AI execution. To run any model entirely on this hardware, the model files must fit completely within this 0.5 GB limit. This page lists the compatible models that can run under these strict hardware constraints.
The quant column indicates the quantization level used to compress the model weights. Quantization reduces the precision of the model parameters to make the file size smaller. For example, Qwen3 0.6B, PixArt-α, PixArt-Σ, and Parakeet TDT 0.6B v2 can fit into the 0.5 GB VRAM when compressed to the Q5_K_M quantization. Smaller models like XTTS v2, Spark-TTS, and CosyVoice 2 can run at Q6_K quantization. Models like VALL-E X, ERNIE 4.5 open weights, F5-TTS, E2-TTS, and ChatTTS require Q8_0 quantization, while StyleTTS 2 fits at FP16 precision.
Running models with a standard 4k context window is extremely difficult on this hardware. As the conversation grows, the context window consumes additional VRAM. This extra memory usage will quickly exceed the 0.5 GB limit of the NVIDIA NVS 310. Users must keep context lengths very short to prevent out of memory errors.
To run larger models, you must use CPU offload. This process splits the model layers between the graphics card and your system memory. We assume your system has 32 GB of system RAM for these calculations. CPU offloading allows you to run models that exceed the physical VRAM of your card, but it comes with a massive performance cost. Processing speeds will drop significantly because system RAM is much slower than video memory.
With CPU offload, you can run Florence-2 base/large or Magenta RT at Q4_K_M quantization. These models need 0.6 GB of VRAM and 2.6 GB of system RAM. Larger models like Stable Diffusion 2.1, Bark, Tortoise TTS, and Riffusion require 0.7 GB of VRAM and 2.7 GB of system RAM at Q4_K_M quantization.
You can also offload Gemma 3 1B, Llama 3.2 1B / 3B, MMS, and CSM-1B. These 1B models require 0.7 GB of VRAM and 2.7 GB of system RAM at Q4_K_M quantization. While offloading makes these models executable, the generation speeds will be very slow.