The "Model may not fit" dialog explained
When you add a GGUF that likely won't run smoothly, Abliteration shows a dialog with the numbers (needed vs available) and lets you cancel or add anyway.
Short answer
When you add a GGUF that likely won’t run smoothly, Abliteration shows a dialog with the numbers (needed vs available) and lets you cancel or add anyway. The warning is honest - it never blocks you from trying.
Detailed explanation
On model add, Abliteration reads the GGUF header, estimates total RAM needed (file size + KV cache at a fixed 8K context estimate + ~1.15 GB baseline overhead), and compares to your free RAM. Three visible outcomes:
Comfortable fit - model needs less than free RAM with headroom to spare. Silent - no dialog, model just adds.
“Tight fit” - model fits but leaves ≤2 GB headroom for other apps. Dialog with a soft warning. Cancel / Add anyway.
“This model fits your PC, but you need to free up RAM” - model needs more than free RAM but fits in total RAM. Dialog explains numbers and suggests closing apps. Cancel / Add anyway.
“This model is bigger than your RAM” - model needs more than total RAM. Strongest warning: may fail to load or run at seconds-per-token. Cancel / Add anyway.
The dialog always shows the cleaned model name (e.g. Qwen3 30B A3B · Q4), not the raw filename, so you know exactly which model it’s talking about.
Why it matters
The dialog exists so you don’t accidentally add a model that will grind your PC to a halt or fail silently. But it’s advisory - Abliteration trusts you to know your machine. If you know you’re about to close Chrome and want to load that 30B, Add anyway works.
The estimate is calibrated but not exact. It uses a fixed 8K KV cache assumption (not the model’s n_ctx_train) - actual RAM at runtime depends on the context you set for the model.
In Abliteration specifically
Triggered on: drag-and-drop of a GGUF onto the Models screen (not the whole app window), or Add Model button. External file drops elsewhere in the app do not trigger the fit check.
The estimate uses the honest KV cache formula (GQA-correct - reads head_count_kv). Fixed context assumption is 8K regardless of what the model was trained for - a 128K-capable model doesn’t need 128K reserved just to load. Actual per-model context is set later in the Edit dialog.
MoE models are estimated by their full file size, not active parameters, because the whole file loads into RAM even though only some experts activate per token.
Related
- RAM requirements - the formula the dialog uses
- Free vs total RAM - the two thresholds that shape which message you see
- Quantization - the fastest way to bring a “won’t fit” model down to “comfortable” is a lighter quant