Qwen3.6 35B A3B quants bite hard

This looks like one of those cases where “fits in VRAM” is not the same as “feels good to use.” The sparse MoE architecture likely makes the active routing paths more sensitive to compression, so quality jumps show up first in agent behavior, not just prose.

• –Qwen’s own model card describes Qwen3.6-35B-A3B as 35B total with 3B activated parameters, and it defaults to thinking mode with tool-use support, which makes any quantization-induced drift more visible in practice.
• –Community reports line up on a simple ladder: q4 is usable but can get loopy or vague, q6 is the likely compromise tier, and q8 is where people start describing a clearly better “feel.”
• –The biggest gains people are noticing are operational, not cosmetic: fewer malformed tool calls, better prompt interpretation, and stronger handling of ambiguous or research-heavy requests.
• –One interesting counter-signal from the thread is that a larger quant can sometimes run faster or more stably than a smaller one once you account for cache behavior, context length, and model-specific quirks.
• –Net: for this model, VRAM saved by going too small may cost more in agent reliability than it looks like on paper.