App vs models
| Component | Typical size |
|---|---|
| Desktop installer | ~10–15 MB |
| Bundled Ollama runtime | ~1–2 GB |
| Default models (14B + 7B) | ~13.5 GB disk |
| Typical first-run total | ~15–20 GB |
RAM tiers
The hub exposes GET /api/system/hardware with your tier and recommended model tags.
| Tier | RAM | Developer primary | Best for |
|---|---|---|---|
| minimal | < 8 GB | llama3.2:3b / qwen2.5-coder:7b | Light chat; cloud hybrid for hard tasks |
| light | 8–15 GB | qwen2.5-coder:7b | Safe local dev |
| recommended | 16–31 GB | qwen2.5-coder:14b | Full software pack + collab |
| heavy | 32 GB+ | 14B + LoRA bases | Multi-model library, CUDA training |
Multi-agent ≠ multi-model
Specialists share one Ollama backend. You are not loading six separate 14B weights. You do pay sequential inference latency and hub memory for repo indexes (capped at 2000 source files).
Hybrid strategy
- Local — iteration, repo Q&A, no per-token bill
- Cloud — hard reviews, long context, when 14B is not enough
- LoRA — disk-efficient specialists on shared bases — see Model App Store
Limits to know
- Installers bundle the runtime, not every model — first pull can be several GB
- Local models vary in collaboration quality — see known issues
- macOS builds are ad-hoc signed — Right-click → Open if Gatekeeper warns
- Browser hub UI is chat-only — use the desktop app for full workspace