Skip to content
niminal

Models and providers

niminal speaks one request shape (OpenAI-style messages and tools) and translates that to each provider’s native API before it goes on the wire.

Wired providers

Section titled “Wired providers”
Provider Default model Key environment variables Endpoint
openrouter openai/gpt-4o-mini OPENROUTER_API_KEY https://openrouter.ai/api/v1/chat/completions
anthropic claude-sonnet-4-6 ANTHROPIC_API_KEY https://api.anthropic.com/v1/messages
google gemini-3.5-flash-lite GEMINI_API_KEY, GOOGLE_API_KEY, or GOOGLE_GENERATIVE_AI_API_KEY https://generativelanguage.googleapis.com/v1beta
openai gpt-5 OPENAI_API_KEY https://api.openai.com/v1/chat/completions
hyper deepseek-v4-flash HYPER_API_KEY https://hyper.charm.land/v1/chat/completions
mistral mistral-vibe-cli-with-tools MISTRAL_API_KEY https://api.mistral.ai/v1/chat/completions
opencode deepseek-v4.1-flash OPENCODE_API_KEY https://opencode.ai/zen/go/v1/chat/completions
opencodezen deepseek-v4-flash OPENCODE_API_KEY https://opencode.ai/zen/v1/chat/completions

OpenRouter is the default at startup. Selecting codex errors: Codex App Server is not wired.

Switch providers with /provider NAME or --provider NAME. Each provider remembers its last model in providers.<name>.last_model inside ~/.niminal/config.json.

/provider alone prints the active name, model, endpoint, and the provider’s standard environment variable. Credentials can also come from ~/.niminal/auth.json or --api-key.

Model selection

Section titled “Model selection”
  • /model alone prints the current id
  • /model ID switches for later turns and saves the choice
  • /model plus Tab opens catalog suggestions filtered from two characters

The catalog comes from models.dev, cached at ~/.niminal/models-dev.json. Use /models refresh to fetch a fresh copy. niminal may also refresh stale cache data in the background at TUI startup.

Thinking levels

Section titled “Thinking levels”

Shared ladder: none, minimal, low, medium, high, xhigh, max.

When you switch models, niminal keeps your saved thinking level and maps it to whatever the new model accepts, using models.dev reasoning metadata plus provider-specific rules.

/thinking with no argument prints the mapped level. Unset thinking leaves the provider default.

Set thinking in the config file, with NIMINAL_THINKING, or with --thinking.

Retries

Section titled “Retries”

Transient model connection failures retry automatically up to three times with exponential backoff. Use /retry after those attempts are exhausted.

Next steps

Section titled “Next steps”