Ampersand v0.1: humans& Trains GLM 5.3 to Collaborate

Ampersand v0.1 is a new AI model from the startup humans& that is trained for one specific job: working well with people and other agents, rather than racing solo through benchmarks. Released on October 20 and available through an Agents API with no waitlist, it is built on GLM 5.3 and comes with an unusually honest model card that lists where it got worse, not just where it got better.

What actually happened

humans& published a blog post and a full model card for Ampersand v0.1. The base is GLM 5.3, a mixture-of-experts model with about 744 billion total parameters, about 40 billion active, and a 1 million token context. Ampersand is text only. Developers reach it with the model ID ampersand-v0.1 through an endpoint that accepts the OpenAI SDK, and it ships with two built-in tools: send_message, to talk to another participant, and pass, to stay quiet when it has nothing useful to add.

That second tool is the whole pitch in miniature. Most chat models answer every turn. This one is trained to decide whether it should speak at all.

How Ampersand v0.1 was trained

The card names two in-house pieces. Persimmon is a user simulator that plays the human side of long conversations. CORO is an off-policy reinforcement learning method used to train on those multi-party sessions. The goal was behavior like sharing information others need, pushing back when a teammate is wrong, and not just agreeing.

The numbers back that up, on humans&'s own tests. Information sharing went from 65% to 89%. Willingness to raise conflict went from 11% to 30%. Sycophancy dropped from 77.3% to 62.6%. On a "Collaboration Over Time" score, Ampersand v0.1 got 12.94 against 9.89 for the base. The company calls its results "tentative", which is a word you rarely see in a launch post.

A model that knows when to say "pass" is rarer than a model that scores one more point on a leaderboard.

Where it got worse

This is the part worth reading. On the SWE-bench Pro Hard-51 subset, Ampersand scored 82.4% against 84.3% for GLM 5.3, so a little coding ability was traded away. On the WMDP cyber benchmark, a test of hazardous security knowledge, the score rose from 80.6% to 89.0%. And the card says prompt-injection susceptibility is "substantially higher than frontier models".

humans& also notes that collaboration costs more. Because the model coaches and checks in, conversations use more turns and more tokens. Pricing is not listed in the card, so we cannot say how that cost lands in dollars.

Why it matters for agent builders

Agent stacks are moving from one bot doing everything to several agents passing work around. That only works if the agents share context and flag mistakes. Most models are tuned to be helpful in a single chat, and CooperBench's 52.5%-to-17.3% collapse for the base model shows how badly that can go when they have to cooperate.

Ampersand is an early attempt at a fix. The base model was not trained for teamwork, and the gap after training is large on humans&'s own tests, though no outside lab has repeated them yet. It is not a frontier model and does not claim to be. It also joins a busy month for niche releases, like the diffusion approach in our Inception Mercury 3 story and the speed play in our MiniMax M3.1 Flash piece.

What to do as a reader

If you build multi-agent systems, try it on a sandbox task where agents have to split work, and compare it with what you use now. Because of the injection warning, do not give it access to wallets, private keys or anything that can move money. Treat any tool output it reads as untrusted.

If you only follow AI news, the takeaway is simple: collaboration is becoming its own benchmark category. Ampersand v0.1 is a small, honest step, not a revolution. Nothing here is investment advice, and there is no token attached to this model, so ignore anyone selling one.

Sources: humans& blog post and the model card.