Usage:
# macOS / Linux (latest release)
curl -fsSL https://kaja.io/setup.sh | bash
# Pin version or repo (put env after the pipe)
curl -fsSL https://kaja.io/setup.sh | REPO=subztep/kaja VERSION=v0.0.1 bash
# Windows (PowerShell)
irm https://kaja.io/setup.ps1 | iex
# e.g. fork: $env:REPO = 'me/kaja'; irm https://kaja.io/setup.ps1 | iex
Model info
parameter_size: 1.7B
Parameters = the “knobs” the model learned during training. More knobs = more knowledge/capability, but also more RAM and slower.
| Size | RAM Needed (rough) | Vibe |
|---|---|---|
| 1B–3B | 1–3 GB | Tiny, fast, dumb-ish. Fine for autocomplete, simple tasks. |
| 7B–8B | 5–8 GB | The “sweet spot” for laptops. |
| 13B–30B | 10–25 GB | Smarter, needs a good GPU or lots of RAM. |
| 70B+ | 40+ GB | Serious hardware territory. |
| 400B+ | Data center | GPT-4 / Claude class. |
So 1.7B = small and fast, runs on pretty much anything, but won’t be winning any IQ contests.
quantization_level: Q8_0
This is a compression trick. Originally each parameter is a 32-bit or 16-bit float (fancy decimal number). Quantization rounds them to smaller integers to save space.
Think JPEG for model weights — lose a little quality, save a lot of size.
Common levels (smaller number = more compressed = worse quality but faster/smaller):
| Quant | Bits/weight | Quality | File size (for 1.7B) |
|---|---|---|---|
| F16 | 16 | Original-ish | ~3.4 GB |
| Q8_0 | 8 | Nearly identical to F16 | ~1.8 GB |
| Q6_K | ~6 | Very good | ~1.4 GB |
| Q5_K_M | ~5 | Good | ~1.2 GB |
| Q4_K_M | ~4 | Decent, most popular | ~1.0 GB |
| Q2_K | ~2 | Noticeably dumber | ~0.7 GB |
Q8_0 is basically “lossless-ish” — about half the size of the original with almost no quality drop. Good default for small models where you’ve got the RAM to spare.
Useful CLI Commands
Check for dependency updates:
bun outdated -r
Update depdendencies for all packages:
bunx npm-check-updates -w -u
Print full output with errors into a file:
bun dev > output.txt 2>&1
Generate Better-Auth secret:
openssl rand -base64 32
Find unused dependencies:
bunx knip
Docker
Start database only:
docker compose up -d db
Connect with any PostgreSQL client to port 5433.