Open Brain
VercelEngineering PostOfficial Source

Ampersand from Humans& now available on AI Gateway

Ampersand is a group-conversation model that chooses whether to message participants or stay quiet after each human message. Its caller-defined tools make that choice an application-controlled action.

Vercel · Nov 6, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Humans& made **Ampersand** available through AI Gateway, free through **November 13, 2026**. Built on Z.ai’s open-weight GLM 5.3, it was trained with reinforcement learning in simulated group conversations lasting up to months of simulated time.

Practical Implication

Unlike a one-to-one chat flow, Ampersand reads messages even when they are not addressed to it, then selects a caller-defined `send_message` or `pass` tool. An application executes the chosen action and records its result.

Agent-Ready Context
Humans& made **Ampersand** available through AI Gateway, free through **November 13, 2026**. Built on Z.ai’s open-weight GLM 5.3, it was trained with reinforcement learning in simulated group conversations lasting up to months of simulated time.

Unlike a one-to-one chat flow, Ampersand reads messages even when they are not addressed to it, then selects a caller-defined `send_message` or `pass` tool. An application executes the chosen action and records its result.

The material describes intended group-conversation behavior but provides no evaluation results. The application must implement both actions; the model’s choice alone does not send a message or control the conversation.
Connected Context · Feed7 Judgment

Ampersand adds a group-conversation model whose intended behavior includes reading messages not addressed to it and choosing whether to send or pass. The application remains responsible for executing and recording that choice, and the brief gives no evaluation results, so availability does not establish conversation quality or reliable control.

Context Map
model#open-models#model-selection
Uncertainty
The material describes intended group-conversation behavior but provides no evaluation results. The application must implement both actions; the model’s choice alone does not send a message or control the conversation.