Kilo Code
Use BlazeInfer models in Kilo Code for VS Code, JetBrains and the terminal.
Kilo Code is an open-source coding agent for VS Code, JetBrains and the terminal. It speaks the OpenAI-compatible API, so BlazeInfer works as a custom provider.
Prerequisites
- A BlazeInfer account with credits and an API key.
BLAZE_API_KEYavailable to the tool.
Configure
Recommended, because tool-calling flags and token limits can only be set in the
file. File: ~/.config/kilo/kilo.jsonc
{
"$schema": "https://app.kilo.ai/config.json",
"model": "blaze/deepseek-ai/deepseek-v4.1-flash",
"provider": {
"blaze": {
"npm": "@ai-sdk/openai-compatible",
"name": "BlazeInfer",
"env": ["BLAZE_API_KEY"],
"options": { "baseURL": "https://blaze.e2enetworks.net/v1" },
"models": {
"deepseek-ai/deepseek-v4.1-flash": {
"name": "DeepSeek V4.1 Flash (BlazeInfer)",
"tool_call": true,
"reasoning": true,
"modalities": { "input": ["text", "image"], "output": ["text"] },
"limit": { "context": 1048576, "output": 16384 }
},
"zai-org/glm-5.3-flash": {
"name": "GLM 5.3 Flash (BlazeInfer)",
"tool_call": true,
"modalities": { "input": ["text"], "output": ["text"] },
"limit": { "context": 1048576, "output": 16384 }
}
}
}
}
}Always set the token limits
Without limit, Kilo sends max_tokens: 32000. That is more than some models
allow, so set limit.context and limit.output in the file config.
Images
Kilo treats a custom model as text-only, so it sends no images unless you declare them. Declare image input explicitly for every model, matching the Vision column on console. Set this line in the model's block. If the block already has the line, replace it:
Vision: Yes
"modalities": { "input": ["text", "image"], "output": ["text"] },Vision: No
"modalities": { "input": ["text"], "output": ["text"] },Never add "image" to a model marked No. If a session starts failing with
"The upstream provider rejected the request." after the agent read an image,
change the model's line to the Vision: No setting, then run kilo -c. If it
still fails, type /undo.
Choose a model
| Model ID | Context | Notes |
|---|---|---|
deepseek-ai/deepseek-v4.1-flash | 1,048,576 | The largest context window, for long coding-agent sessions. Reasoning is on by default and comes back in the reasoning field. Accepts images. Use up to 1,048,576 tokens of context. |
zai-org/glm-5.3-flash | 1,048,576 | A 1,048,576-token context window, for long coding-agent sessions. Supports tool calling. Does not accept images. |
See the console for every model, including ones that are not recommended for coding agents.
Verify it works
Ask Kilo Code to read and edit a file in a small repository. A working setup lists the BlazeInfer model and completes the edit.
Configuration tips
- Set the token limits. If limits are not set, they become 0, which silently turns off automatic context compaction. This is the most important setting on this page.
{env:VAR}is not usable in project config files. Kilo Code rejects the whole project file (ConfigInvalidError, thenModel not found). Use theenvlist shown above instead.- Context size. Models at 32K or below are not recommended for agents.
- Images. A custom model is text-only unless you declare image input. See Images.
Privacy note
Kilo Code has telemetry on by default. To turn it off, add
"experimental": { "openTelemetry": false } to kilo.jsonc.