Models / Qwen
Qwen3-Coder-Next
Strengths
MoE coder built for agentic workflows. 3B active / 80B total. >70% on SWE-Bench Verified with the SWE-Agent scaffold. 256K native context. Apache 2.0.
Weaknesses
80B total weights need ~48GB to fit at Q4. Active-param speed but not active-param memory. Coding-specialized; pick something else for general chat.
Qwen3-Coder-Next is the small-active-param MoE built for production coding agents. With 3B active parameters drawn from an 80B total mixture, it inference-runs like a 3B but reasons like a model 10-20x its active size. Designed for long-horizon agent loops: code edits, tool calls, error recovery, repo-scale planning.
The 256K native context plus broad CLI/IDE scaffold compatibility (Claude Code, Cline, Aider, Continue, etc.) make it the open-weight default for agentic coding in 2026.
When to pick it
- Building a coding agent and you can spend a 48GB GPU on it.
- You want frontier-class agentic code performance under Apache 2.0.
- Multi-step tool-use scenarios where speed-per-token matters and you can afford the weight footprint.
When to skip it
- General chat, reasoning, or non-code tasks. Pick Qwen3-8B or Phi-4 Reasoning instead.
- Your hardware floor is 24GB or smaller. The 80B weights won't fit comfortably below 48GB at usable quantization.