Helper and judge models
The small local models that do one job each, what happens without them, and how to install them.
Besides the chat model, 8gent uses a few local models for specific jobs. All are optional: without them, 8gent carries on and tells you what it is doing instead.
| Job | Model | Size | Without it |
|---|---|---|---|
| System One judge: is this shell command safe? | Selene 1 Mini, 8B | about 5 GB | Commands are checked by the safety rules and the read-only allowlist only. Guarded asks you instead |
| Vision: describe images, PDFs, the screen | qwen2.5vl:7b | about 6 GB | Image description is unavailable |
| Memory search by meaning | nomic-embed-text | about 0.3 GB | Memory search uses full-text search only |
| /goal judge | Any local model other than the one doing the work | - | The run relies on its budget |
| /build planner, engineer, judge | Matched to models you have installed | - | - |
Install them
ollama pull hf.co/AtlaAI/Selene-1-Mini-Llama-3.1-8B-Q4_K_M-GGUF # System One judge
ollama pull qwen2.5vl:7b # vision
ollama pull nomic-embed-text # memory embeddingsThen, for vision, inside 8gent in your project:
/vision model qwen2.5vl:7bMemory use
The System One judge is shared: one copy on your machine serves every tab, agent and terminal. The first time a command needs it, 8gent prints one line naming the model, where it loads and its size.
To keep memory free, skip the judge (System One then uses its rules and allowlist only) or turn System One off with EIGHT_SYSTEM_ONE=0.
In testing
8J, a judge of our own, is being tested against a ladder of real commands. When it passes, it replaces Selene. Until then Selene is the judge.