Get the app

@-mentioning a model

Type @ in the input and pick from the popup.

Short answer

Type @ in the input and pick from the popup. The chosen model’s display name is inserted exactly. In Council mode, @-mentioned models generate the reply; outside Council it’s a hint that has no effect.

Detailed explanation

Triggering the popup. Type @ and a dropdown appears listing all installed models by their display name. As you keep typing, the list filters. Click a model or press Enter to insert its full name.

How the inserted mention looks. When you pick a model from the popup, its display name is inserted with spaces replaced by underscores and the · Q4 quantization suffix dropped. So Llama 3.2 3B · Q4 becomes @Llama_3.2_3B in the input. Behind the scenes, the underscores are normalised back to spaces when matching against your installed models - display name and the token stay in sync.

How matching works. Model names are compared exactly (case-insensitive) after normalizing Unicode variations of hyphens (Windows sometimes replaces - with en-dash – or em-dash — - Abliteration handles this transparently). Underscores and spaces are treated as equivalent, so @Fast_Llama matches Fast Llama.

No silent fallback. If you type a mention that doesn’t match any installed model, Council will show an inline error rather than silently picking the “closest” model. This prevents surprises like @Qwen3-30B accidentally invoking @Qwen 2.5 0.5B.

Visual feedback.

  • Matched @-mentions in the input text turn blue.
  • Unmatched (typo, uninstalled) names stay grey/default text color.
  • Autocomplete inserts the exact tokenised model name - no autocorrect interference from Windows.

Outside Council mode. @-mentions in single-model chats are just text - they have no effect on routing. The currently selected model handles the reply either way. Council mode is what makes @-mentions functional.

Why it matters

Precise mention resolution matters most in Council mode where you’re comparing model behaviors - you need @DeepSeek to actually invoke DeepSeek, not some other model with a similar name. The blue/grey coloring gives instant feedback before you hit Send.

In Abliteration specifically

The autocomplete popup appears above the input pill (not below - it opens upward from the input). Filter is instant as you type. Escape closes the popup without inserting.

Custom-renamed models (via Edit dialog) appear in autocomplete under their custom name, not the original filename. So if you renamed Llama-3.2-3B-Instruct-Q4_K_M to Fast Llama, you’d see Fast Llama in the popup and it inserts as @Fast_Llama.

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