Model Configuration

Cody CLI supports 16 models via the AINative gateway. You can select a model per-session or set a default. The default model is deepseek-4-flash on the free tier and kimi-k2 on paid plans.

Available Models

Model ID	Context	Best For	Plan
`deepseek-4-flash`	128K	Default — coding, tools, reasoning	Paid
`kimi-k2`	128K	Default on paid plans — coding + reasoning	Paid
`kimi-k2.6`	128K	Improved K2 — coding + reasoning	Paid
`kimi-k3`	256K	Flagship reasoning + large context	Paid
`gpt-oss-120b`	64K	Fast coding on Cerebras (2,000+ tok/s)	Free
`glm-5`	128K	Coding + reasoning	Free
`qwen3-coder-next`	256K	Fast coding, large context	Free
`llama-3.3-70b-instruct`	128K	General text	Free
`qwen3-coder-flash`	128K	Fast iteration	Paid
`qwen3.5-397b`	128K	MoE flagship — coding + reasoning	Paid
`qwen-coder-32b`	128K	Code quality	Paid
`claude-sonnet`	200K	Frontier coding + tools	Paid
`claude-opus`	200K	Most capable — complex agentic work	Paid
`gpt-5`	400K	Frontier reasoning, largest context	Paid
`gpt-4o`	128K	Coding + tools	Paid
`o4-mini`	128K	Fast reasoning	Paid

Selecting a Model

Per-session flag

cody --model deepseek-4-flash

Interactive model picker

Run /model inside an interactive session to switch models. The picker is filtered to the models available on your plan.

Set a default in settings

{
  "model": "kimi-k2"
}

Add this to your ~/.cody/settings.json.

Context Limits

Each model has a maximum context window (64K to 400K tokens). Cody automatically manages context to stay within limits, summarizing older turns when needed.

Tool Support

All 16 models in the registry support the full Cody toolset (file read/write, bash, search, MCP tools). Tool-calling reliability is strongest on models tagged tools in the registry — which currently includes every model except llama-3.3-70b-instruct, which is text-focused.