humans& Trains Ampersand to Know When Staying Silent Helps Groups More

AI lab humans& has released Ampersand v0.1, a post-trained version of the 744-billion-parameter GLM 5.3 open-weight model. Ampersand is designed for long-running, multi-person conversations, where it can send a message or explicitly remain silent after each turn. Humans& trained it in multi-agent simulations that unfold over weeks of simulated time.

For developers, Ampersand matters because group-chat agents create a different control problem from single-user assistants. They must track several people, notice scheduling or factual conflicts, share relevant information, and avoid interrupting exchanges that require no help. Humans& trained the model around those collaborative behaviors instead of relying solely on task-completion rewards or short preference comparisons.

Silence becomes an API action

Ampersand’s API makes participation a tool decision. A basic integration follows three steps:

1. Create one session for each conversation.
2. Forward every message with the sender’s name attached.
3. Expose two tools: send_message and pass.

After every incoming message, the model selects one of those tools. A named pass action lets the host application distinguish intentional restraint from a timeout, empty response, or generation failure.

In the published demo, Sam says he will be away Friday and Alex wishes him a good trip. Ampersand passes on both messages. When Maya later proposes a Friday 10 a.m. rehearsal involving all three people, the model flags the scheduling conflict. The exchange illustrates the intended behavior: intervention when shared context makes it useful, followed by silence when it does not.

Long conversations strain standard RL

Ampersand starts with GLM 5.3 and applies additional reinforcement learning for collaboration. Its simulated users come from Persimmon, humans&’s large-scale user model. Persimmon assigns users implicit goals, profiles, and preferences that evolve during an episode. Simulated participants can reject unhelpful suggestions, resist sycophantic responses, and revise their goals or opinions.

Long simulated episodes create an off-policy reinforcement-learning problem because the policy generating experience can fall behind the policy being updated. One episode may span many trainer updates, and a single turn may contain several tool calls. The resulting policy staleness varies across turns, which humans& says made conventional importance-ratio clipping and masking unstable.

Humans& developed CORO, short for Contrastive Off-policy Reward Optimization, to handle that drift. CORO compares trajectories within a batch and derives a relative training signal from those comparisons. This removes the need to estimate an expected value for every simulation seed, an especially difficult calculation when user goals and conversations change over long episodes.

A dynamically adjusted Lagrange multiplier controls KL divergence, a measure of how far the trainer policy has moved from the sampler policy. The multiplier keeps that drift within a fixed budget during training instead of relying on a static penalty coefficient.

Keeping a 744B run moving

• Asynchronous reinforcement learning: Samplers generate conversations on separate devices while trainers update the model, with both sides matched by token throughput.
• Low-precision rollouts: An NVFP4 recipe and custom sparse-attention kernels reduce the compute cost of generating long conversations.
• Live optimization: A meta-optimizer adjusts reward weights and batch composition during the run, allowing the team to respond to reward hacking without restarting the full job.
• Owned hardware: Humans& says training ran on its own hardware in data centers established with DeepInfra.

Coordination gains, coaching costs

Humans& reports results across teamwork, cooperative coding, information sharing, and conflict handling. These figures come from the company’s evaluations.

CooperBench offers the clearest comparison with the base model. GLM 5.3 records a 52.5% success rate when working alone on a repository, then falls to 17.3% when paired with another agent. Ampersand reaches 39.0% in the cooperative setting. The post attributes the improvement to more frequent communication and explicit coordination with its partner.

On the company’s broader simulation evaluation, Ampersand delivers the highest reported value toward users’ implicit goals while incurring more interaction cost than other leading models. That benchmark term is separate from API pricing. Humans& attributes much of the penalty to coaching behavior, including guidance or encouragement when a user wants a direct answer. Such behavior may add friction in products optimized for terse responses.

All benchmark figures are vendor-reported. The release material summarized here does not cite independent replication, confidence intervals, or studies based on live group-chat traffic. The results therefore describe performance under the published test conditions, while production reliability remains unresolved.

The API is live at $1 per million input tokens

At publication, Ampersand is available without a waitlist through an OpenAI-compatible Agents API. Prices are quoted per million tokens:

OpenAI SDK compatibility allows developers to point the official client at the humans& base URL and use the familiar Agents interface. Output billing includes hidden reasoning tokens, so the visible response length may not represent the full output charge.

The API separates reusable agent configuration from conversation state. An agent defines the assistant and its tools. Each session receives a copy of those settings and maintains its own message history, allowing several isolated conversations to share one agent configuration.

The release details summarized here do not specify context-window limits, rate limits, latency targets, data-retention terms, regional hosting options, or uptime commitments. Those operational constraints will affect production deployments involving private team conversations.

The real test starts in group chats

Group chats, project rooms, planning tools, and collaborative coding products are the clearest targets for Ampersand. Humans& demonstrates the model with a replay involving three siblings planning a 40th anniversary over five weeks, a workload that requires memory, selective participation, conflict detection, and coordination across several people.

The broader research claim concerns reward design. A common recent reinforcement-learning recipe uses verifiable rewards for math, code, and tool execution, where success can be checked automatically. Collaborative work has delayed consequences, private preferences, social costs, and goals that change during the interaction. Humans& uses Persimmon to generate those dynamics and CORO to train on the resulting long-horizon episodes at 744-billion-parameter scale.

Live deployments can now test the questions left open by the release: whether Ampersand stays quiet consistently, preserves context across long conversations, handles disagreement without excessive coaching, and improves outcomes for real groups under production latency and cost constraints.