On Oct 20, AI startup humans& released its first model, Ampersand v0.1, pitched as a collaborator that helps groups reach individual and shared goals rather than an assistant that completes tasks alone. It is trained from Z.ai's GLM 5.3 (about 744B total, about 40B active parameters) with long-horizon multi-party RL against simulated users played by humans&'s own user model, Persimmon. It is available now, with no waitlist, through an API compatible with OpenAI's Agents API.
Key Takeaways
- ✓Base: GLM 5.3 MoE with about 744B total and about 40B active parameters and a 1,048,576-token context (official model card)
- ✓Long-term collaboration simulation (vendor figures): 12.94 for Ampersand versus 9.89 for base GLM 5.3
- ✓Collaborative Intelligence teamwork 59.1% (GLM 5.3: 54.3%); CooperBench cooperative success 39.0% versus 17.3% for GLM 5.3
- ✓Trade-off: SWE-bench Pro v2 (Hard-51) 82.4%, slightly below the base model's 84.3%; humans& also admits it sometimes over-coaches
- ✓Access: model ID ampersand-v0.1, usable with the official OpenAI SDK pointed at the humans& base URL
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background
Most LLM training optimizes for finishing tasks and short-term preference scores, which pushes models toward autonomous agents rather than collaborators that fit into a team or group chat. Real collaboration involves unstated goals, conflicting priorities and shifting needs, where knowing when to speak or stay quiet matters as much as what to say. humans&'s first model, Ampersand v0.1, targets that gap.
Architecture
Ampersand is trained from Z.ai's open GLM 5.3 (MoE, about 744B total and 40B active parameters, 1M context) with async RL in multi-party, long-horizon simulations. Simulated users are played by humans&'s own user model, Persimmon, which pushes back, resists sycophancy and changes its mind, with rollouts spanning hundreds of thousands of tokens. To handle heavy off-policy drift, the team introduced CORO (Contrastive Off-policy Reward Optimization), plus an NVFP4 RL recipe and a meta-optimizer that adjusts training as it runs.
Benchmarks
All figures are vendor-reported and marked tentative: 12.94 on the long-term collaboration simulation versus 9.89 for GLM 5.3; 59.1% versus 54.3% on Collaborative Intelligence; on CooperBench, GLM 5.3 drops from 52.5% solo to 17.3% cooperatively while Ampersand reaches 39.0%. The cost is SWE-bench Pro v2 (Hard-51) at 82.4%, slightly under the base model's 84.3%, and humans& says its interaction cost is higher than other leading models, partly from unsolicited coaching.
Getting started
v0.1 is API-only via the Agents API with no waitlist; sign in on the site for a key. The API follows OpenAI's Agents API, so point the official OpenAI SDK at ampersand.humansand.ai/v1 with model ampersand-v0.1. For group chats, open one session per conversation, tag each message with its sender, and give the model send_message and pass tools so it chooses whether to reply. Prices and limits are on the docs' Pricing and limits page.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.