Editing a model's display name and context

Short answer

Click ⋯ on a model in the Models tab → Edit. You can change how the model is named in the UI and how much context it gets. Change the auto-detected chat template if needed.

Detailed explanation

The Edit dialog has three main sections:

Display name - three options:

• Auto-cleaned - the default readable name (e.g. Llama 3.2 3B Instruct · Q4)
• Original filename - the raw GGUF name (Llama-3.2-3B-Instruct-Q4_K_M)
• Custom - free text field for whatever you want (My abliterated tester, Fast responder, etc.)

The choice affects everywhere the model appears: Models list, chat model selector, @-mention autocomplete, drawer chat titles, fit-check dialog.

Context - two modes:

• Auto (recommended) - Abliteration picks a sensible context based on your free RAM and the model’s trained ceiling. Usually 4K-16K depending on the situation.
• Manual - a slider stepping through discrete powers of two: 2048, 4096, 8192, 16384, 32768, 65536, 131072. Minimum 2048. Maximum is min(n_ctx_train, 131072). Live RAM cost shows in the label as you drag; a red warning appears if the setting would need more RAM than you have free.

Below the slider, a line notes what the model was trained for: “Model was trained for up to N tokens”.

Chat template - a 16-option dropdown (Auto-detect · ChatML · Qwen · Llama · Mistral · DeepSeek · Gemma · Phi · GLM · Kimi · Command-R · Yi · InternLM · Baichuan · OLMo · Falcon). Default is Auto-detect, which uses the template embedded in the GGUF. Change this to a specific family if the model produces garbled output (see Weird output). Manual override wins over auto-detection.

Save - the Save button commits changes. Closing the dialog without Save discards.

Why change context manually:

• You know you’ll be working with long documents (bump to 32K+)
• You want to save RAM for other apps (drop to 4K)
• The model’s default context is too aggressive for your machine

Why it matters

Custom names help if you have many similar models - call them “Llama fast”, “Llama smart”, “Qwen tester” instead of decoding version numbers every time. Manual context lets you optimize the RAM/capability trade-off per model instead of using one-size-fits-all. Manual chat template is the fix for garbled output.

In Abliteration specifically

Path: Models tab → click ⋯ on any model → Edit. Only the ⋯ button opens the menu - no right-click handler.

Per-model settings persist to manifest.json in the models folder (not in the SQLite database - the DB only holds chats and messages). Reinstalling the app keeps your per-model settings if you keep the models folder.

Context changes take effect on the next model load - chats already open with that model keep their current context until you reload.

The slider’s live RAM number uses the KV cache formula, which is GQA-correct (reads head_count_kv from the GGUF header). It does not shrink for MoE (KV cache does not benefit from MoE routing - only tok/s does).

Related

• Model naming - how the auto-cleaned name is generated
• Context slider - deeper on how the slider works
• Context length - the concept behind the slider
• Chat template auto-detect - how template detection works and the 16-option dropdown