Hermes Agent Users Hit Q2 Wall

The hot take is that this reads less like a Hermes problem than a "the model was quantized too hard" problem.

• –Qwen3.5-35B-A3B Q2_K sits in the "very low quality but surprisingly usable" tier, so weak output is expected rather than surprising.
• –On a 24GB RTX 3090, a higher-quality setup like Qwen3.5-27B at Q4_K_M or a better 35B quant will usually feel much more coherent than squeezing for minimum VRAM.
• –Hermes Agent adds orchestration, memory, and tools, but it cannot recover reasoning quality the base model no longer has.
• –For mixed chat, research, and agentic work, consistency usually beats raw parameter count; a dense mid-size model can feel better than a heavily compressed MoE.