Alibaba Qwen3-Coder-Next: Open Agentic Coding Model
Alibaba's Qwen team released Qwen3-Coder-Next on 2 February 2026, describing it as an open-weight model designed specifically for coding agents and local development that uses a hybrid-attention MoE architecture to deliver strong coding and agentic capabilities at significantly lower inference cost
Tool Overview
Features, steps and FAQ below
Features
- ✓ Built for coding agents and local development: Qwen says it is an open-weight language model designed specifically for coding agents and local development
- ✓ Hybrid-attention MoE architecture: Qwen says it is built on Qwen3-Next-80B-A3B-Base with a novel hybrid-attention and MoE architecture, with 80B total parameters and only 3B activated per token
- ✓ Agentically trained at scale: Qwen says it was agentically trained at scale on large-scale executable task synthesis, environment interaction and reinforcement learning
- ✓ Significantly lower inference cost: Qwen says it obtains strong coding and agentic capabilities with significantly lower inference costs
- ✓ 256K context and tool calling: Qwen says the default context length is 256K and that it excels at tool calling, working with coding agents such as Claude Code and Cline
How to Use
- Get the model weights and docs from the official Qwen blog, GitHub or Hugging Face
- Deploy locally or on a server with vLLM, SGLang or Ollama
- Set a suitable context length (256K by default, or smaller as needed) and configure the tool-call parser
- Plug it into coding agents such as Claude Code, Cline or OpenClaw for local development
FAQ
What is Qwen3-Coder-Next?
An open-weight language model released by Alibaba's Qwen team on 2 February 2026, which Qwen says is designed specifically for coding agents and local development.
What is special about its architecture?
Qwen says it is built on Qwen3-Next-80B-A3B-Base with a novel hybrid-attention and MoE architecture, with 80B total parameters and only 3B activated per token, significantly lowering inference cost.
How long is its context window?
The official docs say the default context length is 256K, and you can reduce it to a smaller value such as 32768 if the server fails to start.
Can it run locally?
Qwen describes it as an open-weight model designed for local development, deployable locally or self-hosted and supported by inference frameworks such as vLLM and SGLang.
Is it free?
Qwen describes it as an open-weight model; its Hugging Face model card lists an Apache 2.0 license, so after downloading the weights local inference costs no per-request fee, and it is also available via hosted APIs such as Alibaba Cloud Model Studio.