Get the app

Council mode - talking to several models

Toggle Council on at the top of a chat, then @-mention models by name to get answers from each one in sequence.

Short answer

Toggle Council on at the top of a chat, then @-mention models by name to get answers from each one in sequence. Great for comparing perspectives; slower than single-model chat because models take turns loading.

Detailed explanation

Turning it on. The Council toggle sits at the top of every chat. Once on, this chat becomes a Council chat until you turn it off. Council state is remembered per-chat - re-opening the chat later restores the toggle position. First time you enable Council, a brief one-time intro appears.

@-mentioning models. Type @ in the input and an autocomplete popup shows all your installed models. Pick one (or several - @Llama @Qwen @DeepSeek in the same message). Matched @-mentions turn blue in the input text; unmatched names (typos, models you don’t have) stay grey.

How it responds. Each mentioned model gets loaded in turn, generates its answer with a badge showing its name, then unloads before the next model loads. All models see the same pre-turn snapshot of the chat history - meaning within a single turn, later models do NOT see earlier models’ just-generated replies (those show in the UI, but aren’t fed forward). Across subsequent turns, prior replies are in history normally.

Sequential loading. Only one model is in RAM at a time. This keeps memory usage predictable but adds a load pause between models (2-10 seconds each). A three-model Council answer might take 30 seconds total. See How Council manages RAM for the reasoning.

Model badges. In Council mode, each assistant reply shows a small badge above it with the model’s display name (verbatim, mixed case as configured - not force-uppercased). Regular single-model chats don’t show badges.

Drawer marker. Chats that have ever used Council mode get a C marker to the left of the chat title. If the chat has also used a reasoning model, the marker becomes Rᶜ - one glyph, superscript c on a full-size R. See Drawer navigation.

No conversational routing. Models don’t know which model wrote earlier turns - they just see assistant: .... If you want them to react to each other by name, mention it explicitly (“respond to what Qwen just said in the previous turn”).

Why it matters

Council is the killer feature for research. Getting several different models’ takes on the same question in one place is genuinely different from asking each separately - you see disagreement patterns, complementary insights, model biases. Especially useful when studying model behavior (abliteration research, alignment differences).

In Abliteration specifically

Toggle location: top of chat, above the message list.

If a model fails to load or errors mid-generation in Council mode, an inline error card appears where the response would have been: ⚠ {Model Display Name} couldn't respond - {reason text}. Other mentioned models continue normally.

If you @-mention a model name that doesn’t match any installed model exactly, an inline error card appears: Model @X not found. Available: @A, @B, @C... (up to 6 available names listed). No silent fallback to the “closest match” - if you meant @Qwen3-30B-A3B-Q4_K_M, that exact name (space-to-underscore normalised) must resolve.

Context compression (if the chat is long enough to trigger it) runs once per turn, shared across all Council models - not per model. See orchestrator_summarization for the real mechanics.

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