Qwen3.5 Debate Weighs Bits, Brawn

My read: for coding and tool use, precision usually matters more once quantization gets extreme, so the smaller q8 model is the safer default unless the larger model is still in a tolerable low-bit regime.

• –Qwen3.5-122B-A10B is a much bigger MoE model on paper, but a 2-bit-class conversion can wipe out the advantage you were buying with scale.
• –Coding and tool calling are brittle tasks: they depend on exact syntax, stable instruction following, and clean function-call formatting, which low-bit quantization tends to hurt first.
• –A 35B q8_0 model usually gives you a more reliable baseline for repo edits, JSON/tool schemas, and code generation than a far larger model at ultra-low precision.
• –The bigger model starts making sense again if the quantization is gentler, or if your workload is more about broad reasoning than precise output formatting.
• –The Kimi K2.6 angle is the same tradeoff in a different costume: once you go to 1-bit territory, you are mostly testing compression tolerance, not getting the full value of the model.