Serving Qwen3.6-35B-A3B With vLLM and Building a Coding Agent With Tool Calling

The article discusses serving Alibaba's Qwen3.6-35B-A3B model locally using vLLM, calling it from Python with the OpenAI SDK, and wiring up tool calling to enable the model to act as a coding agent.

Details

Alibaba's Qwen team released the Qwen3.6-35B-A3B model on April 16, 2026 under the Apache 2.0 license. It is a sparse mixture-of-experts (MoE) model with 35 billion total parameters but only about 3 billion active per token. The model scores 73.4% on the SWE-bench Verified benchmark and 37.0 on the MCPMark, making it one of the strongest open-weight models for agentic coding tasks. The article provides instructions for serving the Qwen3.6-35B-A3B model locally using the vLLM framework. vLLM version 0.19.0 or later is required, as older versions do not support the Qwen3.6 MoE architecture. The article covers basic inference-only setup as well as enabling tool calling, which allows the model to act as a coding agent that can call external tools and services. The article also demonstrates how to call the Qwen3.6-35B-A3B model from Python using the OpenAI SDK, with the local vLLM server acting as the API endpoint. The 'reasoning-parser' and 'tool-call-parser' flags enable the model to generate internal reasoning steps and call external tools as part of its response, improving its capabilities for coding tasks.