Dense or MoE: which open model to pick
Four open-weight models, hardware requirements that match reality, and a decision tree that does not lie about VRAM. As of May 2026.
Filed under infrastructure
The coding model that stays on your servers (2026)
Qwen3-Coder-30B-A3B activates 3.3B parameters per token and fits a 48 GB workstation at AWQ 4-bit. Apache 2.0 licence.