Ki Complete
Code suggestions as you type, from KvcHub's own code model running on your computer. Your code never leaves your machine, it works offline, and it's free.
- Tab accepts a suggestion. Cmd+Right (Ctrl+Right on Windows and Linux) accepts one word. Esc dismisses it. Alt+\ asks for one.
- The Ki item in the status bar turns suggestions on or off, switches models, restarts, and opens the log.
What it does well
- Finishes the line you're typing, then stops. In the middle of a statement you get the rest of that line. When a line opens a block (ends with
:, {, (, [ or =>) you get the block. On a new line you get the next line or block.
- Stays out of the way. No suggestions in the middle of existing code, none that repeat code already below the cursor, none that repeat the line above.
- Knows your project. It shows the model pieces of files you recently opened or edited, so suggestions use your own names.
- Easy to recognise. Ki's suggestions are green, with a faint green background, instead of grey (turn off with the Green Suggestions setting).
- One autocomplete at a time. In Antigravity, Ki Complete turns off Antigravity's own autocomplete once, with a button to turn it back on.
- Never writes out keys. It won't suggest an API key, token or private key that isn't already in your file.
- Quick. While you type, a suggestion typically takes 0.1-0.13 seconds on an Apple M5 Pro, because the model server reuses its work between keystrokes, and typing the start of a suggestion keeps the rest without asking again.
Models
| Model |
Size |
Notes |
| Ki Complete (default) |
1.6 GB, about 2 GB memory |
KvcHub's model, trained for exactly this |
| Larger: 7B |
8.1 GB download |
general code model, not tuned by KvcHub, slower |
| Largest: 30B |
32.5 GB download |
needs 64 GB of memory |
You can also use any GGUF code model with fill-in-the-middle support that you already have: status bar, Model, "A model file I already have".
How good is it
Measured on questions it never trained on (details in the Ki Complete project's README):
|
Before training (Qwen2.5-Coder 1.5B) |
Ki Complete |
| Real code: the rest of the line exactly right |
31-40% |
36-44% |
| ... with other files as context |
34-41% |
39-45% |
| HumanEval-Infilling, one missing line: tests pass |
88.4% |
89.4% |
| HumanEval-Infilling, several missing lines: tests pass |
50.8% |
53.4% |
| Model answers padded with lines nobody asked for (before clean-up) |
65% |
3% |
| Typical wait on real code (first request at a spot) |
0.59 s |
0.23 s |
Requirements
- llama.cpp, which runs the model. On a Mac:
brew install llama.cpp
- About 2 GB of free memory for Ki Complete.
Ki Complete starts the model server itself on port 8012 and stops it when you close the editor. To use a llama.cpp server you already run, set Ki Complete: Server Url.
Settings
Search the settings for "Ki Complete": model, model file, languages to stay quiet in (plain text and Markdown by default), how long to wait after you stop typing (80 ms), the most lines per suggestion (16), whether to use other files, and more.
Privacy
Everything runs on your computer. Ki Complete sends nothing anywhere: no telemetry, no accounts. The only network request it can make is downloading a model you choose, from Hugging Face.
Credits and licenses
Ki Complete is licensed under Apache-2.0. The Ki Complete model is a fine-tune of Qwen2.5-Coder-1.5B (Apache-2.0, by the Qwen team); the larger models are Qwen2.5-Coder-7B and Qwen3-Coder-30B-A3B (Apache-2.0). Models run with llama.cpp (MIT).