LocalLLaMA crowns Qwen3.6 over GLM-4.7

The blunt read is that your hardware is already competitive; the gap is mostly model choice and workflow, not a bad machine. The thread’s center of gravity has shifted toward Qwen3.6 27B as the practical local coding pick, while larger preview models are treated as interesting but awkward to run interactively.

• –Commenters say Qwen3.6 27B is the current sweet spot for a 5090 because it fits the VRAM budget and stays fast enough for interactive coding
• –GLM-4.7-Flash is being treated like last cycle’s answer, with newer mentions including DeepSeek V4 Flash preview and MiniMax M2.7
• –The recurring advice is to use local models for planning, refactors, and code search, while keeping hosted models for the hardest write-heavy tasks
• –DGX Spark is framed more as a coherent-memory desktop appliance than a straight upgrade for code quality, so it is hard to justify over a strong consumer GPU unless you need that specific form factor