Why does my 30B MoE model get a "small model" warning?

Short answer

It doesn’t get the “small model” warning - MoE models get a different warning that specifically explains the MoE trade-off. If you’re seeing the small-model warning on a 30B file, that’s a bug (please report).

Detailed explanation

Two different warnings, different triggers:

Warning 1 - Small dense model (⚠ Small model (<3B)): Fires only on dense models under 3B parameters. Explains they may hallucinate on factual questions.

Warning 2 - MoE model (⚠ MoE model: XB total, only ~YB active per token): Fires on any MoE model (any size). Explains that answers may be shallower than a dense model of equivalent total size, though speed matches the active-parameter equivalent.

Example for Qwen3-30B-A3B:

⚠ MoE model: 30B total, only ~2B active per token - answers may be shallower than a dense 30B model. Speed matches ~2B though.

This is not the small-model warning - it’s a specific MoE explainer that acknowledges both the value (fast) and the limitation (potentially less deep reasoning than a dense 30B).

Why the confusion: A MoE 30B like Qwen3-30B-A3B has ~2B active parameters. If Abliteration used active params to decide “is this small?”, it would incorrectly flag the 30B MoE as small. Detection uses total params for the small-model rule - a 30B MoE doesn’t qualify.

The MoE warning is separate and specifically explains the situation without misleading language.

Why it matters

Warnings should describe reality accurately. “Small model may hallucinate” on a 30B file would be confusing - the 30B stores 30B worth of knowledge, even if only some of it activates per token. The MoE warning explains the actual trade-off.

In Abliteration specifically

MoE detection reads expert_count from the GGUF header. If present and > 1, model is treated as MoE. Total and active parameter counts inform the warning text.

If for some reason a MoE model gets the small-model warning instead of the MoE warning (or both), that’s a detection bug - report it with the model name so the MoE detection can be checked against that GGUF’s header layout.

Related

• MoE vs dense - the underlying architecture
• Small model hallucinations - the other warning