Pick a model (or set up BYOK)
How to choose between auto, the Anthropic ladder, a BYOK provider, or a free local model.
AI
Preview — early access.
Overview
The IDE's agent panel can route to a lot of models. Most of the time auto is the right call — this guide is for the times you want to deliberately pick something else, and how to get a BYOK or local model actually working.
How it works
Three questions decide it: Do you want Creytix to decide? Leave it on auto — it already picks Haiku for trivial fixes, Sonnet for standard work, and Fable for hard multi-file work. Do you want a specific paid provider you already have a key for? Add that provider's key to the BYOK vault and pick its model by name. Do you want it free and local? Pick one of the local Creytix models — no key needed at all. They run on the Creytix Model Server, which ships its own engine and loads the weights already on your machine.
Steps
- •
Default to auto
For most day-to-day edits, leave the model picker on auto and let the router match the model to the task.
- •
Want a specific cloud model instead?
Check the Models tab for the exact id (e.g.
grok-4,kimi-k2,gemini-flash,gpt-4o-mini), then add that provider's key under BYOK key vault if it isn't already configured. - •
Want it free, with nothing to configure?
Make sure the Creytix Model Server is running (
npm run creytix:models:serve, or install it as a login service withnpm run creytix:models:install-launchd), then pick a local model (e.g.qwen2.5-coder-7b,llama3.1-8b,gemma2-9b) — it needs no vault key. - •
Confirm the key status
Before relying on a BYOK model, check its dot in the Models tab: ● configured, ○ needs a key.
- •
Match the model to the task
Cheap/local and parallel-only models are best for cross-checks and bulk drafts; reserve the primary-eligible Anthropic tiers (Sonnet, Opus, Fable) for the change you're actually going to apply.
Capabilities
- Auto ladder — Haiku (trivial), Sonnet (standard), Fable (hard) picked automatically from your instruction.
- BYOK cloud lane — Grok 4, Kimi K2, Gemini Flash/Flash-Lite, Qwen via DashScope, GPT-4o mini, Codestral, Command R, DeepSeek Chat, plus Groq/Cerebras/Together/Fireworks/OpenRouter hosted models.
- Free local lane — Qwen2.5-Coder, Llama 3.1/3.3, Gemma2, DeepSeek-Coder-V2, Phi-4, Mistral and more, served locally by the Creytix Model Server, keyless.
- Trust labeling — the Models tab marks each model primary-eligible, parallel-only, or candidate, so you know which are meant for the main edit versus a cross-check.
Limits & honest scope
- A BYOK model with no key configured won't respond — it fails informatively (shown as "needs a key") rather than silently.
- Local models need the Creytix Model Server running; it provisions its own engine and loads the weights already on disk, so there's nothing separate to install or pull.
- Not every model is trusted for the primary edit — several are parallel-only or candidate-trust, meant for fan-out drafts and cross-checks, not the change you actually ship.