Run huge AI models on the PC you already have.

The biggest open models, like DeepSeek V4 and Qwen3, are mixtures of experts. They're split into thousands of small expert networks, and each word only uses a few of them. So the whole model never has to sit in fast memory at once.

Pelica keeps the parts every word needs on the GPU, along with the experts the model reaches for most. The next tier waits in RAM. Everything else stays on your SSD and is read just in time: Pelica runs the next layer's router early to guess which experts are coming, and gets the top pick right about 99% of the time.

Experts move up as they're used and drop down when they go quiet, so the fast memory keeps working on whatever your conversation needs right now. None of this changes the model's answers. Every speedup is checked to produce exactly the same words.

Measured on a mid-range gaming PC: Ryzen 5 8400F, 32 GB of DDR5, an RTX 5060 Ti with 16 GB, and one PCIe 4 NVMe drive.

A token is a word or part of one. For comparison, Colibrì runs GLM-5.3-Flash (320B) at 0.76 tokens/s on the same PC; Pelica runs it at 3.2.

# the engine also works on its own, from a terminal
pelica --model Qwen3-30B-A3B-Q4_K_M.gguf -p "Why is the sky blue?"

Pelica is an app, not a service. Models run on your computer, your chats are saved on your computer, and there's no Pelica server, account system or analytics. It only goes online when you ask it to: downloading a model, searching the web, checking for updates, or syncing through your own Google Drive if you turn that on.

All of the code is on GitHub, so you can check this for yourself. The privacy policy has the details.