Qwen3-Coder-Next too big for 16GB

The subtext is simple: local agentic coding is still a hardware budgeting game, and the “best” model on paper is often not the one you can keep loaded all day.

• –Qwen3-Coder-Next is the buzzed-about name here, but the official Qwen family still points at much larger checkpoints and long-context tooling, so it is not the easy 16GB answer.
• –The practical shortlist in the thread is exactly what you would expect for homelab work: smaller quantized models that can follow instructions reliably without blowing VRAM.
• –Commenters also steer the stack discussion toward llama.cpp over Ollama, which matters as much as model choice once you are squeezing every token/sec out of a consumer card.
• –For Docker Compose and NixOS migration help, instruction-following and tool use matter more than leaderboard bragging rights.