Models (Neural Net)¶
Engineering → Neural Net → Models decides which model does each job. Everything runs in Ollama on the Starbase.
This PC¶
Scan this PC reads the graphics card, memory and processor, and recommends a model for each job. Nothing is downloaded until you say so. Use every recommendation picks them all.
Chat size¶
The personas share one brain, in three sizes for different graphics cards (Automatic picks from the scan):
| Size | Brain | Graphics memory | Download |
|---|---|---|---|
| Standard | Llama 3.3 8B (Q6_K) | about 10.5 GB | 6.6 GB |
| Lite | Llama 3.3 8B (Q4_K_M) | about 6 GB | 4.9 GB |
| Ultralite | Llama 3.2 3B | about 3.2 GB | 2.2 GB |
Build all makes the persona models that are missing in the chosen size (the first downloads the shared brain).
Rebuild all makes the installed ones again from Starfarer's Modelfiles (after their notes change); the ones you
had are kept as copies (<name>-old-<date>). A persona's name opens its card in Personas.
A model that's missing in the chosen size is said plainly ("Nova can't answer yet: its model isn't installed"), never quietly swapped for another.
Jobs¶
| Job | Used for |
|---|---|
| Chat | each persona's own model |
| Deep | thinking things through; away missions, when chosen |
| Vision | looking at pictures |
| Code | writing scripts and code |
| Reading | reading long pages and documents in parts |
| Embed | finding things it remembers by meaning (past conversations) |
Every job but Chat uses only the model picked here; None turns that job off and the chat AI does what it can itself. Changes apply on the next reply. A tag says whether each is ready, in memory, or needs a download.
Downloading and memory¶
Download a model takes any Ollama model name ("qwen3:14b"). In memory now shows what's loaded and how much graphics memory it takes. Starbase loads and unloads models as jobs need them; with Share the GPU on (Network) it only unloads models it loaded itself, so other AI apps, games and ComfyUI are safe.