Chat — one question, a council of models

Open Chat →

Grafy Chat is a multi-model surface: pick from local and cloud models, stream answers with visible reasoning, then put models side by side on the same question — and let a data fuser read every column and write one combined answer.

Walkthrough: question to fused verdict

A clean slate

The composer with its Think and Tools switches, plus mic dictation; Sessions keeps every conversation.

Choose your model

A searchable catalogue with maker, context window, and Free / Vision pills. Local models cost nothing.

Answers stream — reasoning included

With Think on, replies carry a collapsible Reasoning fold; markdown renders fully and every turn records its model.

A streamed answer with reasoning fold and a table

Put models side by side

Add a model as a column: one question goes to the whole party, each column answers in parallel under a shared context cap.

Two model columns answering the same question

Let the fuser write the verdict

The data fuser reads every column and streams one synthesis, crediting each contributor.

The data fuser panel with a fused answer

Generate images without leaving the thread

Switch the composer's tool to Image and describe the picture — it lands inline as an assistant turn, saved with the conversation.

An image generated inline in a chat conversation

Going deeper

The model catalog

The picker lists what the server actually serves: local models (free, private, always available) and cloud models routed through hosted providers and billed on real usage — the same credit balance as every other Grafy generation. Each entry shows its maker, context length, and capability pills; Vision models accept image attachments, and the search box filters the whole catalog as you type. Switching the model mid-conversation applies from the next turn, and each reply is stamped with the model that wrote it.

Think mode & reasoning

Think asks the model to reason before answering. The chain of thought streams into a separate, collapsible fold above the reply — readable when you want to audit the logic, out of the way when you don't. Answers render full markdown (tables, code with highlighting, lists), a failed turn offers a one-click Retry that re-runs just that panel, and replies can be read aloud with the built-in text-to-speech. The mic button dictates into the composer.

Compare columns & the data fuser

The + tab turns a chat into a party: each added model gets its own column with its own transcript, and every question you send goes to all of them at once. The header shows the shared context window — capped to the smallest model in the party — and Grid / Single switches between seeing everything and focusing one column. Close a column to stop asking that model; a joining model is caught up on the conversation so far.

Turn on the Data fuser and pick a judge model: once every column answers, the fuser reads them all and streams one clearer, combined answer into its own panel, with a FUSED FROM chip row crediting each contributor. The panel folds three ways — split, chats-only, fuser-only — and the fusion is saved with the conversation like any other transcript.

Images in chat

Two directions: attach an image (vision models will actually look at it — screenshots, whiteboards, photos), or generate one with the Image tool — the composer placeholder flips to “Describe your image…”, and the result arrives inline, produced by the same image engine Creator uses and saved in the conversation history.