Qwen3-Coder-Next: Efficient 80B Model with 3B Active Params for Coding

Overview

Qwen3-Coder-Next is a new open-weight LM positioned for coding agents and local development. The announcement highlights a mix of agent-focused training data, an efficiency-oriented parameter design, and broad tooling compatibility aimed at developer workflows.

Agentic training and benchmarks

The model is reported to have been trained with 800K verifiable tasks plus executable environments, an emphasis that targets agentic workflows where code must be executed or validated as part of the training loop. On agent-centric evaluation, the release cites SWE-Bench Verified performance of over 70% when paired with the SWE-Agent scaffold. The team also notes competitive results on SWE-Bench Pro while maintaining a compact active footprint.

Efficiency and architecture

A central technical claim is an efficiency–performance tradeoff: the model has 80B total params while operating with a much smaller 3B active parameter footprint during inference. That design aims to reduce runtime resource needs while preserving results comparable to larger open-source models on agent benchmarks. No additional architectural specifics are provided in the announcement.

Ecosystem and integrations

Qwen3-Coder-Next is presented as supporting multiple code-focused runtimes and frameworks, including OpenClaw, Qwen Code, Claude Code, and several web- and browser-oriented development uses. The list also mentions Cline among supported tools. Release materials link to hosted model repositories and supplementary documentation for those interested in hands-on use.

Where to find more

Repository and documentation links provided with the announcement:

Further details and the original announcement are available at the source: https://x.com/i/status/2018718453570707465