Qwen3.5-27B sparks 70B model hunt

This reads like a calibration problem, not just a size problem: the poster wants better dot-connecting, and Qwen3.5’s own docs show it already leans on thinking mode and long context. The cleanest path is a larger, well-tuned cloud model or a higher Qwen3.5 tier, not a random "uncensored" fine-tune that may trade away the reasoning they’re after.

• –Official Qwen3.5 docs say the model defaults to thinking mode and supports 262K native context, so prompt structure and context hygiene matter a lot.
• –The Qwen3.5 ladder already climbs to 122B-A10B and 397B-A17B, and those bigger tiers do better on knowledge, reasoning, and agent/search benchmarks than 27B.
• –The user's caution about derestricted models is fair: removing safeguards can also remove calibration and instruction-following quality.
• –Their feed -> question -> scaffold -> search -> second opinion workflow is basically the right way to squeeze nuanced answers out of any cloud LLM.