Model selectors

Every model in ka is addressed by a selector:

vendor/model@effort

• vendor (the provider): anthropic, openai, google, zai, deepseek, groq, openrouter, ollama, lmstudio, llamacpp, vllm, … plus 500+ catalog models across 27 seeded vendors.
• model: the model id, verbatim, including any colons (e.g. qwen3.5:9b).
• effort: optional reasoning effort (low / medium / high, some models add max). The separator is @ precisely because model ids may contain colons.

ka --model anthropic/claude-sonnet-5@high
ka --model ollama/qwen3.5:9b

Where selectors are set

Priority from lowest to highest; each layer overrides the last:

1. [roles] in ka.toml (default, fast)
2. model in ~/.config/ka/ka.toml
3. model in .ka/ka.toml (trusted projects)
4. KA_MODEL environment variable
5. --model flag

ka.toml

model = "ollama/qwen3.5:9b"
effort = "high"

[roles]
default = "anthropic/claude-sonnet-5"
fast    = "groq/llama-3.3-70b"     # cheap lookups, subagents

The /model picker

/model with no argument opens a picker over the entire catalog:

• type-to-filter across vendors and model ids
• per-row key/env status, context window, and price (only for verified prices; see below), plus each model’s reasoning levels where the dialect exposes them
• Enter on an unmatched filter sets it as a custom vendor/model selector
• picking a model persists it as your default for future conversations

/settings edits model, mode, and effort together; press s to persist the whole panel. /thinking [level] picks the reasoning level directly (Enter persists; the named form is session-scoped). The thinking level always shows beside model and mode whenever the model exposes reasoning control.

Roles

Roles let you keep one main model and a cheap one for subagent work (pathfinder, fast lookups). [roles] fast is consulted where a cheap model is appropriate; default is your main.