Aider

Settings#

export OPENAI_API_BASE=https://api.ai.itsincom.org/v1
export OPENAI_API_KEY="sk-its-…"

aider --model openai/qwen3.8-27b

The openai/ prefix tells Aider to use the OpenAI-compatible API at OPENAI_API_BASE. The model id after it is the one in the catalogue.

The context window and the edit format#

Aider does not know ITS INCOM AI's models, and without help it warns that it does not know their context window. Two files in the root of your repository (or in your home directory) fix that.

.aider.model.metadata.json:

{
  "openai/qwen3.8-27b": {
    "max_input_tokens": 131072,
    "max_output_tokens": 32768,
    "litellm_provider": "openai",
    "mode": "chat"
  }
}

.aider.model.settings.yml:

- name: openai/qwen3.8-27b
  edit_format: diff
  use_repo_map: true

For another model, add an entry to each file with its id and the context window from Which model. On openai/qwen3.8-27b, max_input_tokens is 131072, and accepts_settings: ["reasoning_effort"] in its settings lets --reasoning-effort none switch its reasoning off.

With edit_format: diff the model sends the lines it changes; with whole it sends every file it edits in full.

What we ran#

In a new repository with one file, hello.py, and the two files above written for openai/qwen3-coder-30b (context window 262144):

aider --model openai/qwen3-coder-30b --yes-always --no-stream \
  --map-tokens 0 --message "Add a function add(a, b) to hello.py that returns a + b." hello.py

What came back, shortened:

Model: openai/qwen3-coder-30b with diff edit format
hello.py
<<<<<<< SEARCH
def greet(name):
    return f"Hello, {name}!"
=======
def greet(name):
    return f"Hello, {name}!"

def add(a, b):
    return a + b
>>>>>>> REPLACE
Tokens: 2.4k sent, 72 received.
Applied edit to hello.py
Commit a76db4f feat: add add(a, b) function to hello.py

Aider read the context window from the metadata file, applied the edit and committed it with a message written by the same model.

Which model#

qwen3.8-27b, with 131,072 tokens of context. Aider sends only the files you add to the chat. qwen3-coder-30b answers fast, for everyday edits. qwen3.8-27b has 131,072 tokens of context and reasons before it answers. With --reasoning-effort none the same kind of edit, adding sub(a, b), came back on 11 October 2026 with 81 tokens received. The others are in Which model.

Limits#

• 32,768 output tokens per answer. With edit_format: diff that is rarely reached; with whole a large file can be cut.
• Commit messages are written by the same model, one more request per commit, unless you name another with --weak-model.