Model Switching

Fermi lets you switch between models at any point during a session. Thinking level is configured as part of the model switch flow.

The /model Command

Type /model during a session to open a hierarchical picker:

/model

The picker shows all configured providers and their models. Select one and the agent switches immediately for the remainder of the session.

For managed providers (Kimi, MiniMax, GLM, DeepSeek, Xiaomi, Qwen), if you select a model whose key is missing, Fermi can prompt you to import a detected external env var or paste the key directly.

Managing API Keys

The /model flow prompts for a key only when one is missing. To manage keys directly — add, replace, remove, or import a provider’s API key at any time — use /key:

/key

It opens a picker of provider endpoints; choose one and then pick an action. Keys live in ~/.fermi/.env (0600 permissions). This is the easiest way to rotate a key or fix one that was entered incorrectly without re-running fermi init.

The /model command itself is useful for:

• Starting with a fast/cheap model for exploration, then switching to a stronger model for implementation
• Moving to a cheaper model when the task becomes routine
• Testing how different models handle the same context

Thinking Levels

After switching models, Fermi prompts you to select a thinking level (if the model supports multiple levels). The available levels vary by provider:

Provider	Levels
Anthropic (Opus 4.7)	off, low, medium, high, xhigh, max
Anthropic (Opus 4.6)	off, low, medium, high, max
Anthropic (Sonnet 4.6)	off, low, medium, high
Anthropic (Haiku 4.5)	off, low, medium, high
OpenAI	none, low, medium, high, xhigh
DeepSeek	off, high, max
Kimi	off, on
GLM	off, on
Xiaomi (MiMo)	off, on
Qwen	off, on
MiniMax	on (M3 also supports off)
GitHub Copilot	follows the underlying model’s levels

A few newer models differ from their provider’s general pattern: GLM-5.2 uses high / max, Kimi K2.7 Code is on only, and Kimi K2 Instruct has no thinking control. The picker always shows the exact levels each model supports.

Higher reasoning depth produces more thorough analysis but uses more tokens and takes longer.

Model Tiers for Sub-Agents

Use /tier to configure which models sub-agents use at different capability levels:

/tier

This opens a picker where you assign specific models to three tiers:

Tier	Typical use
high	Complex reasoning, architectural decisions
medium	Standard implementation work
low	Simple tasks — file listing, grep, basic edits

When the agent spawns a sub-agent with model_level="low", it uses the model assigned to the low tier. This saves cost on routine work while keeping the main agent on a powerful model.

Adding More Models

Models come from the providers you configure during fermi init. To add more:

1. Re-run fermi init to add new providers
2. For local providers, start the server and re-run init to discover new models
3. For OpenRouter, any model available through their API can be used

For OpenAI (ChatGPT Login), use fermi oauth or /codex to authenticate. For GitHub Copilot, use /copilot to log in first.