Qwen3.5 users want 18B middle ground

Open-weight model families are now being judged as much by hardware fit as by benchmark scores, and Qwen3.5's lineup is exposing that tradeoff in real time.

• –Several commenters say the 9B model is impressively close to 27B on some benchmarks, but still too weak for users who want a reliable local coding or general-purpose daily driver.
• –The 27B dense model is treated as the quality pick, yet many users see it as too slow or too memory-hungry for comfortable local use on consumer machines.
• –Qwen3.5-35B-A3B is repeatedly suggested as the compromise option, but the thread makes clear many users do not see a small MoE as a full substitute for a true mid-size dense model.
• –The discussion reflects a broader local AI trend: parameter counts only matter insofar as they map cleanly onto common RAM and VRAM tiers like 16GB, 32GB, and 64GB systems.