> ## Documentation Index
> Fetch the complete documentation index at: https://qitor.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Model family matrix

> The built-in QitOS v0.4 family presets and their default harness policies.

| Family    | Transport         | Default protocol       | Fallback chain                                          | Tool delivery      | Context window | Recommended models                             |
| --------- | ----------------- | ---------------------- | ------------------------------------------------------- | ------------------ | -------------- | ---------------------------------------------- |
| Qwen      | OpenAI-compatible | `json_decision_v1`     | `xml_decision_v1 → react_text_v1`                       | `api_parameter`    | 128k           | `Qwen/Qwen3-8B`, `qwen-plus`, `Qwen/Qwen3-32B` |
| Kimi      | OpenAI-compatible | `json_decision_v1`     | `react_text_v1`                                         | `api_parameter`    | 128k           | `kimi-k2-0905-preview`, `moonshot-v1-128k`     |
| MiniMax   | OpenAI-compatible | `minimax_tool_call_v1` | `terminus_xml_v1 → terminus_json_v1 → json_decision_v1` | `api_parameter`    | 128k           | `MiniMax-M2.5`                                 |
| `gpt-oss` | OpenAI-compatible | `json_decision_v1`     | `react_text_v1`                                         | `api_parameter`    | 128k           | `gpt-oss-120b`, `gpt-oss-20b`                  |
| Gemma 4   | OpenAI-compatible | `json_decision_v1`     | `react_text_v1`                                         | `prompt_injection` | 128k           | `gemma-4-31b-it`, `gemma-4-26b-a4b-it`         |
| OpenAI    | OpenAI-compatible | `json_decision_v1`     | `react_text_v1`                                         | `api_parameter`    | 128k           | `gpt-4.1`, `gpt-4o`                            |
| Anthropic | OpenAI-compatible | `react_text_v1`        | `json_decision_v1`                                      | `prompt_injection` | 200k           | `claude-sonnet-4-5`                            |
| Gemini    | OpenAI-compatible | `xml_decision_v1`      | `json_decision_v1 → react_text_v1`                      | `prompt_injection` | 1M             | `gemini-2.5-pro`                               |
| DeepSeek  | OpenAI-compatible | `json_decision_v1`     | `tool_use_xml_v1 → react_text_v1`                       | `api_parameter`    | 128k           | `ds-v4-pro`, `ds-v4-flash`, `deepseek-chat`    |
| GLM       | OpenAI-compatible | `json_decision_v1`     | `xml_decision_v1 → tool_use_xml_v1 → react_text_v1`     | `api_parameter`    | 128k           | `glm5.1-w4a8-4maas`, `glm-4`                   |

## Gold presets

The first five families (Qwen, Kimi, MiniMax, `gpt-oss`, Gemma 4) are the **gold presets** — the most thoroughly tested configurations that the QitOS team validates on every release. When you are getting started or running a benchmark comparison, prefer a gold preset.

The remaining five (OpenAI, Anthropic, Gemini, DeepSeek, GLM) are **compatibility presets**. They provide correct protocol and context defaults for their families but may evolve as dedicated adapters are added.

## Use the matrix correctly

This table is not a ranking.

It shows the default harness (the wiring layer that connects a transport, parser, and protocol) contract QitOS will use when you ask for a family preset (a reusable configuration bundle for a model family).

If you need to override one part, do it explicitly:

* override the protocol with `build_model_for_preset(protocol=...)`
* override the tool delivery with `build_model_for_preset(tool_delivery=...)`
* extend a preset with `preset.override()` — see [Family presets](/concepts/family-presets)

Do not silently fork the whole example.