> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.mixlayer.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.mixlayer.com/_mcp/server.

# Qwen 3.5/3.6

> Available Qwen models, thinking modes, and recommended sampling settings

Qwen is Mixlayer's general-purpose model family. It spans small, low-latency models through large mixture-of-experts models for difficult reasoning and agentic work.

## Model details

* **Context window:** 131K tokens
* **Capabilities:** Text · Vision · Reasoning · Tools

## Available models

| Model                                                                                    | Identifier               |
| ---------------------------------------------------------------------------------------- | ------------------------ |
| **Qwen 3.5 4B (Free)** Free prototyping and short, low-stakes tasks.                     | `qwen/qwen3.5-4b-free`   |
| **Qwen 3.5 9B** High-volume chat, classification, and summarization.                     | `qwen/qwen3.5-9b`        |
| **Qwen 3.6 27B** General-purpose reasoning and instruction following.                    | `qwen/qwen3.6-27b`       |
| **Qwen 3.6 35B (3B active)** Stronger reasoning, coding, and agents with MoE efficiency. | `qwen/qwen3.6-35b-a3b`   |
| **Qwen 3.5 397B (17B active)** Hard reasoning, multi-step coding, and agentic workloads. | `qwen/qwen3.5-397b-a17b` |

## Thinking modes

Qwen supports a **thinking mode**, with reasoning returned in `reasoning_content`, and a **non-thinking mode** for faster direct answers. See [Reasoning](/reasoning) for request examples.

## Mixlayer platform defaults

When these parameters are omitted from a request, Mixlayer applies the following defaults:

| Identifier               | Default settings                                                    |
| ------------------------ | ------------------------------------------------------------------- |
| `qwen/qwen3.5-4b-free`   | `temperature=1.0`, `top_p=0.95`, `top_k=20`, `presence_penalty=1.5` |
| `qwen/qwen3.5-9b`        | `temperature=1.0`, `top_p=0.95`, `top_k=20`, `presence_penalty=1.5` |
| `qwen/qwen3.6-27b`       | `temperature=1.0`, `top_p=0.95`, `top_k=20`                         |
| `qwen/qwen3.6-35b-a3b`   | `temperature=1.0`, `top_p=0.95`, `top_k=20`                         |
| `qwen/qwen3.5-397b-a17b` | `temperature=0.6`, `top_p=0.95`, `top_k=20`                         |

## Recommended sampling

| Mode                      | Recommended sampling                                                                          |
| ------------------------- | --------------------------------------------------------------------------------------------- |
| Thinking — general tasks  | `temperature=1.0`, `top_p=0.95`, `top_k=20`, `presence_penalty=0.0`, `repetition_penalty=1.0` |
| Thinking — precise coding | `temperature=0.6`, `top_p=0.95`, `top_k=20`, `presence_penalty=0.0`, `repetition_penalty=1.0` |
| Non-thinking              | `temperature=0.7`, `top_p=0.80`, `top_k=20`, `presence_penalty=1.5`, `repetition_penalty=1.0` |

> **Info**
>
> Qwen also recommends `min_p=0.0`. Mixlayer does not currently expose
> `min_p`, so omit it and use the gateway default.

For the mixture-of-experts variants, active parameters primarily govern inference cost and latency while total parameters contribute to capability.