Skip to content
Neural Daily

ModelsNews

humans& Ampersand: a model trained to work with groups

humans& has released Ampersand v0.1, a model trained to help groups of people reach their goals rather than replace their work. It is built on GLM 5.3 and is available now through an Agents API with no waitlist.

By: Mikola Máté Patrik · · 2 min read

humans& model release card: Ampersand, a collaborator model designed to help groups of people achieve their goals
Image: humans& · Press kit · Official blog social image (og:image) · Source

What it is

humans& argues that current training optimizes for task completion and shallow preferences, which pushes models toward autonomy that replaces work. Ampersand was instead trained in long, multi-party simulations. It tracks what each person wants, when it is worth speaking up and when it is better to stay quiet.

Numbers from the company

  • Collaborative Intelligence Benchmark: 59.1% at high effort, against 49.8% for the best open-weight public result. Versus its GLM 5.3 base, information sharing rose from 65% to 89% and conflict handling from 11% to 30%.
  • CooperBench (agents working in the same repository): 39.0% cooperative success. GLM 5.3 drops from 52.5% solo to 17.3% when cooperating.
  • In simulations spanning months, Ampersand delivered the most value toward users' unstated goals. The company admits it also racks up more interaction cost, because it sometimes coaches people who just want an answer.

These are the company's own measurements, and the main evaluation is modeled on a product it has not released yet. There is no independent check so far.

How it was trained

With simulated users driven by a separate user model, Persimmon. They push back, resist sycophancy and change their goals along the way. Training used async RL with a new off-policy algorithm, CORO. A meta-optimizer flags reward hacking during training so the run can be corrected without a restart. The team used an NVFP4 RL recipe on its own hardware and in datacenters built with DeepInfra.

For developers

The API follows OpenAI's Agents API, so the official OpenAI SDK works by changing the base URL (API reference). For group chat the model gets two tools, send_message and pass, and after each message it chooses to reply or stay quiet. Pricing per million tokens is $1 for input, $0.20 for cached input and $5.20 for output, reasoning tokens included.

With group bots in Slack or Teams, the problem is often not what the bot says but how much it talks. A model trained to stay quiet as well addresses exactly that. It is quick to try in an internal channel.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles