humans&’s launch artwork for Ampersand, its first model (illustrative). Image: humans&
humans&, the AI lab that raised $480 million from backers including Nvidia and Jeff Bezos before it had shipped anything, has released its first model, and it is built on a Chinese one. Ampersand v0.1 is trained from GLM 5.3, the open-weight model released by Beijing's Z.ai (Zhipu), and is available now to developers through an API.
The startup calls Ampersand "the first model trained for people to achieve more together": instead of finishing tasks on its own, it is meant to sit in group chats and workplace threads, work out what each person is trying to do, and decide when to speak and when to stay quiet.
A US lab on a Chinese base model #
humans& launched on January 20 as a "human-centric frontier AI lab", with founders from xAI, Anthropic, Google DeepMind, OpenAI, Meta, Stanford and MIT. Its seed round, led by SV Angel and co-founder Georges Harik, valued it at $4.48 billion, and GV, Emerson Collective and Marissa Mayer are among the other investors it lists.
Nine months on, its first product is not a model trained from scratch. The model card lists GLM 5.3 as the base: a mixture-of-experts model with about 744 billion parameters (about 40 billion active per token) and a context window of just over a million tokens. Ampersand takes text in and gives text out.
The lab is open about it. "We trained Ampersand from GLM 5.3, released by Z.ai, and we thank them for their contribution to the open-weight community," the launch post says. It is a sign of how far Chinese open models have come that a richly funded US lab picked one as its starting point, at a time when US rivals such as Nvidia-backed Reflection are pitching their own open models as the Western alternative.
Trained on simulated groups of people #
Ampersand was trained with reinforcement learning in long, multi-person simulations, some covering months of simulated time. The simulated users come from Persimmon, a separate humans& model built to behave like real people: they push back, resist flattery and change their minds. The model is scored on how much it moved each person towards their long-term goals, minus the cost of interactions they didn't want.
In practice, developers give it two tools, one to send a message and one to "pass". In the launch post's example, a colleague mentions he is away on Friday; later, another suggests a Friday rehearsal with everyone, and Ampersand steps in, unprompted, to point out the clash.
The company's own scores #
All of the results humans& reports come from its own runs, and the model card warns they are "tentative" and "may change as runs finish". On its headline simulation test, it says Ampersand gives the most value towards users' unstated goals of any model it tried, but admits it also runs up a higher "interaction cost" than other leading models, largely because it offers coaching or encouragement when people "simply want an answer".
It reports 59.1% on the Collaborative Intelligence teamwork benchmark, against 49.8% for the best open-weight result on the public board, and 39.0% success on CooperBench when two agents code together, up from 17.3% for GLM 5.3. It scores slightly below its base model on a hard software engineering set (82.4% against 84.3%).
The safety table is more mixed. Accuracy on WMDP's cybersecurity questions, a proxy for hazardous hacking knowledge, rose from 80.6% to 89.0%, and the card says prompt-injection success remains "substantially higher than frontier models". That matters for a base model that Anthropic recently said can build working cyber exploits on its own. On the other side, refusals of harmful requests improved, and conversations with sycophancy on its internal test fell from 77.3% to 62.6%. humans& says keeping it API-only lets it monitor use and respond to abuse.
Price and availability #
There is no waitlist: developers sign in at ampersand.humansand.ai for a key, and the API follows OpenAI's Agents API, so the OpenAI SDK works with a changed base URL. Pricing is $1 per million input tokens ($0.20 cached) and $5.20 per million output tokens (US dollars). The terms require users to be 18 or over and bar use where the law forbids it, but the launch materials don't list supported countries, and the weights have not been released. The lab says a product it built alongside the model will launch "soon".
Why it matters #
humans& is betting that the next leap in AI is models that work well with groups of people, not agents that replace them, and it has billions in valuation riding on that idea. That its first model leans on a Chinese base, and picks up stronger cyber knowledge along the way, shows how much US startups now depend on open weights from China.
Sources: humans&: Introducing Ampersand; Ampersand v0.1 model card; Ampersand pricing and limits; humans& launch page; Bloomberg.