Get the app

@-mention doesn't recognize my model

The name must match exactly (case-insensitive).

Short answer

The name must match exactly (case-insensitive). Use the autocomplete popup to insert it - don’t type manually. If it stays grey, the name doesn’t match any installed model.

Detailed explanation

Symptoms:

  • @ModelName stays grey in the input instead of turning blue
  • Council mode returns an inline error “Model @X not found” after sending
  • Autocomplete popup doesn’t show your model

Common causes and fixes:

1. Wrong exact name. Model names are the display name (auto-cleaned), not the original filename. Qwen2.5-7B-Instruct-Q4_K_M won’t match - it’s stored in Abliteration as Qwen 2.5 7B Instruct · Q4.

Fix: type @, look at the autocomplete popup, pick from there. Guarantees exact match.

2. Hyphen substitution by Windows. Windows can auto-replace ASCII hyphen - with typographic dashes: en-dash – (U+2013), em-dash — (U+2014), non-breaking hyphen ‑ (U+2011). Looks similar visually but is a different character to the matcher.

Fix: use the autocomplete popup. It inserts the exact stored characters. Never type model names manually if they contain hyphens.

3. Autocorrect / spellcheck interference. Windows text prediction can silently rewrite what you type in a TextBox. Doesn’t matter for autocomplete-inserted names (those bypass autocorrect) but can trip manual typing.

Fix: same - use the autocomplete popup.

4. Model doesn’t exist. You may have deleted it, or you’re thinking of one from a different app/session.

Fix: check Settings → Models. If it’s not listed, it’s not installed.

5. Custom-renamed model. If you renamed a model via ⋯ → Edit → Custom, the new name is what the @-mention needs, not the auto-cleaned original.

Fix: check the model’s current name in Settings → Models. Use that exact name in @-mention (or via autocomplete).

Why it matters

@-mentions in Council mode are precise on purpose - no silent fallback to the “closest match” model, so you never get a different model’s answer than the one you asked for. The trade-off is that mentions must be exact. The autocomplete popup exists to make this painless.

In Abliteration specifically

Unicode normalization is applied to both the typed mention and stored model names before matching. So typographic dashes vs ASCII hyphens shouldn’t be a problem - but relying on that is risky. The autocomplete route is reliable.

If a model name really won’t match no matter what you do, the fallback is renaming it via ⋯ → Edit → Custom to something short and typeable (e.g. “llama-fast”). Then @llama-fast matches easily.

Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Get the app Browse the catalog
Open in Abliteration