packages/kilo-docs/pages/code-with-ai/agents/model-selection.md
Here's the honest truth about AI model recommendations: by the time I write them down, they're probably already outdated. New models drop every few weeks, existing ones get updated, prices shift, and yesterday's champion becomes today's budget option.
Instead of maintaining a static list that's perpetually behind, we built something better — a real-time leaderboard showing which models Kilo Code users are actually having success with right now.
👉 See what's working today at kilo.ai/models
This isn't benchmarks from some lab. It's real usage data from developers like you, updated continuously. You'll see which models people are choosing for different tasks, what's delivering results, and how the landscape is shifting in real-time.
While the specifics change constantly, some principles stay consistent:
{% tabs %} {% tab label="VSCode" %}
/models to open the model picker./variant to open the reasoning effort selector.kilo.jsonc config file.{% /tab %} {% tab label="CLI" %}
Ctrl+X m or /models) to switch models.--model flag to kilo run (e.g., kilo run --model claude-sonnet-4-20250514).model key in kilo.jsonc, or configure per-agent models in the agent section.--model flag → Per-agent config → Last used in session → Global config → Recent models → First available.{% /tab %} {% tab label="VSCode (Legacy)" %}
{% /tab %} {% /tabs %}
For complex coding tasks: Premium models (Claude Sonnet/Opus, GPT-5 class, Gemini Pro) typically handle nuanced requirements, large refactors, and architectural decisions better.
For everyday coding: Mid-tier models often provide the best balance of speed, cost, and quality. They're fast enough to keep your flow state intact and capable enough for most tasks.
For budget-conscious work: Newer efficient models keep surprising us with price-to-performance ratios. DeepSeek, Qwen, and similar models can handle more than you'd expect.
For local/private work: Ollama and LM Studio let you run models locally. The tradeoff is usually speed and capability for privacy and zero API costs.
Using an unlisted model? You can register any model — including fine-tunes, newly released models, or custom local models — by adding it to your config file. See Custom Models for details.
One thing that doesn't change: context window size matters for your workflow.
Check our provider docs for specific context limits on each model.
{% callout type="tip" %}
Be thoughtful about Max Tokens settings for thinking models. Every token you allocate to output takes away from space available to store conversation history. Consider only using high Max Tokens / Max Thinking Tokens settings with modes like Architect and Debug, and keeping Code mode at 16k max tokens or less.
{% /callout %}
{% callout type="tip" %}
Recover from context limit errors: If you hit the input length and max tokens exceed context limit error, you can recover by deleting a message, rolling back to a previous checkpoint, or switching over to a model with a long context window like Gemini for a message.
{% /callout %}
When an agent delegates work to a subagent (via the task tool), the subagent inherits the parent agent's model by default. You can override this per subagent in your config:
{% tabs %} {% tab label="CLI" %}
{
"agent": {
"explore": {
"model": "anthropic/claude-haiku-4-20250514"
}
}
}
This sets the explore subagent to always use Haiku regardless of the parent's model. Any subagent without a model override uses whatever model the invoking agent is running.
{% /tab %} {% tab label="VSCode" %}
Subagents inherit the model currently active in the primary agent session — the model shown in the selector at the bottom of the chat. To bypass inheritance and pin a specific model for a subagent:
kilo.jsonc:{
"agent": {
"explore": {
"model": "anthropic/claude-haiku-4-5"
}
}
}
The Settings UI writes the same agent.<name>.model entry, so either method produces the same override. Subagents without an explicit model continue to inherit whatever the invoking agent is running.
{% /tab %} {% tab label="VSCode (Legacy)" %}
In the legacy extension, each mode has Sticky Models — switching from one mode to another (e.g., Code → Architect) uses whatever model you last selected for that mode, not the model from the mode you came from. This means you can assign different models to different modes:
The model selection is remembered per mode across sessions.
{% /tab %} {% /tabs %}
For details on configuring subagent models, see Custom Subagents.
The AI model space moves fast. Bookmark kilo.ai/models and check back when you're evaluating options. What's best today might not be best next month — and that's actually exciting.