Introducing ZixcAI: An Agentic AI Chat Interface Powered by a 770B Parameter Model (80B Active)

Today, the author launches ZixcAI, an agentic AI chat interface built on a custom 770 billion parameter mixture-of-experts model with 80 billion active parameters per token. Unlike most AI chat products released in 2026, which are wrappers around existing APIs, ZixcAI is entirely new. The interface, the agent runtime, and the model routing are all created by the author.

ZixcAI is an agentic AI assistant, meaning it can do more than just chat. When you send a message, the model can decide to execute code in a per-user sandbox, read and write files in an isolated workspace, fetch URLs from the public web, browse the web, analyze images via a dedicated vision sub-agent, remember information across sessions, quote, bookmark, and branch conversations.

The interface is a dark, minimal, keyboard-first chat app, but the backend is an agent loop with retries, circuit breakers, and per-turn audit logs.

The core of ZixcAI is a mixture-of-experts transformer with 770B total parameters and 80B active parameters per token. Rather than using a dense model, an MoE architecture is used, which gives the same compute cost as a much smaller dense model. This allows ZixcAI to have 10x the total knowledge of a dense 80B model, while paying for memory (all experts must be loaded) but not compute on every token.

ZixcAI responds with streamed token-by-token output via Server-Sent Events. Reasoning is displayed in a separate "thinking" panel, and tool calls appear inline as animated cards with status indicators. When you send an image to ZixcAI, it routes it to a vision sub-agent, which returns a detailed text description that is then fed into the main reasoning model, allowing the model to respond using the combined context. The main model remains text-only for speed and cost, but image understanding is handled where it matters.

ZixcAI remembers facts about you across conversations, has persistent memory that you can edit, and supports conversation branching. Each message has a "fork" button, allowing you to create a new conversation starting from that message without affecting the main thread. The main model is a text-only text model, but ZixcAI can understand images by routing them to a separate vision sub-agent and feeding the returned text description into the main model.