Alibaba's Qwen3-Coder-Next Beats Models 10x Bigger on Coding Agent Tasks
Alibaba's Qwen team ships an 80B mixture-of-experts coder that activates just 3B parameters yet matches Claude Sonnet on agentic coding tasks.
Takeaways
- Qwen3-Coder-Next is an 80B MoE model that activates only 3B parameters per token
- Scores 70.6 on SWE-bench Verified and 44.3 on SWE-bench Pro, matching models 10-20x larger
- Priced at $0.12/$0.80 per million input/output tokens on OpenRouter
- Native 256K context, extendable to 1M via Yarn, supports 358 programming languages
- Available on Hugging Face with FP8 and GGUF quantizations
- Works with Qwen Code, Claude Code, Cline, and browser-use agent scaffolds
Alibaba's Qwen team has released Qwen3-Coder-Next, an open-weight coding model built around a specific question: how small can you shrink the compute footprint of an agentic coding model before performance falls off a cliff? Based on the published benchmarks, much smaller than most people assumed. The model uses a sparse MoE design with 80B total parameters and only 3B activated per token, yet posts numbers comparable to models running 10 to 20x more active compute.
An 80B model that runs like a 3B one
The core architectural bet is a hybrid attention plus mixture-of-experts stack inherited from Qwen3-Next-80B-A3B-Base. During inference, only 3 billion of the 80 billion parameters activate on any given token. To push capability at that footprint, the team leaned on agentic training: large-scale synthesis of verifiable coding tasks paired with executable environments, so the model learns directly from environment feedback through mid-training and reinforcement learning.
Rather than throwing more parameters at coding, they generated huge numbers of tasks that could actually be executed and checked, then let the model practice against those environments. The reward signal comes from whether the code runs and passes checks, not from human preference alone.
The full Qwen3-Coder family now covers three sizes:
- Qwen3-Coder-480B-A35B-Instruct: the flagship, 480B total parameters with 35B active per forward pass (8 of 160 experts)
- Qwen3-Coder-30B-A3B-Instruct: a lighter self-hosting option
- Qwen3-Coder-Next: the new 80B/3B-active hybrid
FP8 and GGUF variants ship for each, so single-GPU and llama.cpp deployments are covered out of the box.