Without the RTX 5090, I can still run a local "AI girlfriend" on the GTX 3050.

After viewing an AI girlfriend Agent model shared by the Chinese community, I tried to rebuild a lighter version to check whether a popular machine can run it or not.

Unlike the original model, which requires about 15GB of VRAM and is usually run on RTX 5080 or 5090, the version I tested has reduced the model size and temporarily removed the voice input to better suit the GTX 3050.

The point I find most interesting is that the entire system runslocally on the machine:

✅ No need to connect API connection

✅ Chat content does not have to be sent to external services

✅ Can customize personality and response

✅ Does not depend on the Internet after preparing enough models

Currently the test version mainly focuses on conversation and voice response capabilities. Due to the small model used, the speed and intelligence certainly cannot be compared to systems running on high-end GPUs, but the results are still quite good for a GTX 3050.

If given additional permissions and connected to other tools, this type of Agent can develop beyond chatting, such as controlling smart devices, opening applications or supporting some daily tasks.

Of course, this is just a technical test, not a complete “AI lover”.😆 But the fact that a personal computer can run LLM, speech recognition and speech generation on its own without the need for an API shows that local AI is becoming more and more accessible.

Who knows, in a few years, the first person will ask you:

“How was your day?”

is not the lover, but the computer placed on the table. 😂

Configuration I use:

  • GPU:GTX 3050
  • LLM:Llama 3.2 3B
  • Voice recognition:Whisper Small
  • AI voice:Qwen3-TTS 0.6B
  • How to communicate:text input instead of live voice chat

#LocalAI #AIAgent #Llama32 #Whisper #Qwen3TTS #GTX3050 #ArtificialIntelligence

Need a reliable technology partner for your business? See DUDI's corporate website design.