humans& Boba and Ampersand v0.1: A Group-Chat AI Built on GLM 5.3 (744B)

humans& has put a product and a model into public use at the same time. Boba, a free assistant you text or drop into a group chat, went live on October 22, 2026. The model behind it, Ampersand v0.1, is open to developers through an OpenAI-compatible Agents API at $1 per million input tokens. The base is Z.ai's open-weight GLM 5.3, which has 744 billion parameters. That matters because some headlines round the size to 700B, and because the interesting claim here is not scale but behavior: Ampersand is trained to decide when to stay quiet.

This post sets out what humans& and its launch coverage actually say, what the numbers mean, and which parts nobody has checked yet. Every performance figure below is vendor-reported unless stated.

Quick reference: what was launched

What Boba actually does

Boba's own site pitches "social intelligence for your goals and your people." You text it from your phone, add it to a group chat, or use it on the web, and it costs nothing with no card. The examples on the page are everyday: a trip chat where it researches prices and routes and splits the cost, reminders and calendar invites, and a family chat that keeps a shared list.

It can connect to email and calendar, and the site says it books things once you say yes. The page also warns that Boba can make mistakes and that messages may be used to improve it. If you are deciding whether to add it to a family chat, that second line is the one to read twice: other people in the chat are sharing their messages with a model that may learn from them.

The product category is not new. Proactive assistants are a crowded idea, and we looked at one recently in Hark Pro versus Muse Dots. What is different about Boba is that the model underneath was trained specifically for multi-person conversations rather than adapted from a single-user chat model by a system prompt.

Why a group chat is a different problem

A single-user assistant answers when spoken to. A group chat breaks that rule in three ways:

1. Most messages are not for the assistant. In a family chat, "have a good trip" is social. An assistant that replies to everything becomes noise within an hour.
2. Context is spread across people. One person says they are away Friday; another proposes a Friday rehearsal two days later. Useful help means connecting the two.
3. Goals are private and drift. People hold implicit preferences, change their minds and push back on suggestions they dislike.

humans& argues that standard training recipes do not teach this. Verifiable rewards work for math, code and tool execution because success can be checked automatically. Collaboration has delayed consequences and social costs. So the company trained Ampersand in long simulations instead.

How Ampersand was trained

A ball following a reward track, illustrating humans& Boba Ampersand reinforcement learning from simulated feedback

According to the AlphaSignal write-up of the model release, Ampersand starts from GLM 5.3 and adds reinforcement learning for collaboration. We covered the base model's launch and benchmarks in GLM-5.3 launch and cyber-defense benchmarks and its open-weights timing in the GLM-5.3 open-weights delay.

The ingredients, as described:

• Simulated users. Simulated people come from Persimmon, humans&'s large-scale user model. Each is assigned implicit goals, a profile and preferences that evolve during an episode. They can reject unhelpful suggestions, resist sycophantic replies and revise their goals.
• Long episodes. Simulations unfold over weeks of simulated time. Launch coverage describes the method as Reinforcement Learning from Simulated Feedback, run over simulated days and simulated people; those figures describe training, not real-world reliability.
• A new algorithm called CORO. Contrastive Off-policy Reward Optimization. Long episodes mean the policy that generated an experience can lag the policy being updated, and humans& says conventional importance-ratio clipping became unstable. CORO compares trajectories within a batch to derive a relative signal, avoiding an expected-value estimate for every simulation seed. A dynamically adjusted Lagrange multiplier keeps drift between sampler and trainer within a fixed KL budget.
• Infrastructure for a 744B run. Asynchronous RL with samplers and trainers on separate devices, an NVFP4 recipe with custom sparse-attention kernels for cheaper rollouts, and a meta-optimizer that adjusts reward weights mid-run to respond to reward hacking. Training ran on humans&'s own hardware in data centers established with DeepInfra.

Readers who follow post-training will recognize the pattern of moving beyond verifiable rewards. For other recent examples of RL recipes on open bases, see Castform's Neon RL post-training and self-play with verifiable rewards. The difference here is that the reward signal comes from simulated people, which is both the novelty and the main risk, since a model trained against simulated users can only be as good as the simulation.

Silence as an API action

Three speech bubbles converging, showing a group-chat model choosing when to speak or pass

The most concrete developer feature is how the API treats restraint. A basic integration, per the write-up, has three steps: create one session per conversation, forward every message with the sender's name attached, and expose two tools, send_message and pass.

After each incoming message the model picks one of the two. Because pass is a named action, your application can distinguish intentional restraint from a timeout, an empty response or a generation failure. That is a small design choice with real consequences for reliability monitoring: you can log pass rates, alert on a sudden shift and test whether the model stays quiet when it should.

The API also separates an agent, which defines the assistant and its tools, from sessions, which each copy those settings and keep their own message history. Several isolated conversations can therefore share one agent configuration. Because it is OpenAI-compatible, you point the official client at the humans& base URL.

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 with all three, it flags the conflict. A longer replay shows three siblings planning a 40th anniversary over five weeks.

The numbers, and how far to trust them

CooperBench gives the cleanest comparison with the base model. GLM 5.3 solves 52.5 percent of repository tasks alone but falls to 17.3 percent when paired with another agent. Ampersand reaches 39.0 percent in that cooperative setting, which humans& attributes to more frequent communication and explicit coordination. Notice that 39.0 is still below the solo figure for the base model, so cooperation is not yet free.

On the company's broader simulation evaluation, Ampersand delivers the highest reported value toward users' implicit goals but at a higher interaction cost than other leading models. Humans& attributes much of that penalty to coaching: guidance or encouragement when a user simply wants a direct answer. In a product built for terse replies, that friction could be a real complaint.

Claimed versus verified

We could not open humans&'s own site from our environment, so the model details here rest on the AlphaSignal write-up and Boba's product page. Treat them as secondary until you confirm against the company's announcement and API docs.

What developers should check first

If you are tempted to build on Ampersand, test these before committing:

1. Pass rate on your traffic. Replay real chat logs, with consent, and measure how often it passes when a human would have.
2. Coaching. Measure how often it adds encouragement or guidance no one asked for. The vendor's own results flag this as a cost.
3. Reasoning-token billing. Output at $5.20 per million includes hidden reasoning tokens, so visible reply length understates the bill. Compare cost per conversation, not per token, against alternatives such as the open models in our GLM-5.3 and Kimi K3 comparison.
4. Privacy terms. Private team or family chats are sensitive. Retention and training use are not specified in the coverage, and Boba's site says messages may be used to improve it.
5. Memory over weeks. The sibling-anniversary demo spans five weeks; your deployment needs the same continuity, and so a session store you control.

Why it matters

Two things stand out. First, a lab used a strong open base, GLM 5.3, and put its differentiation entirely in post-training and a product. That is the same economics behind several recent releases, and it means a post-train can be shipped quickly on someone else's weights. Second, restraint is becoming a first-class capability. Most assistants are rewarded for responding; treating "say nothing" as an explicit, loggable action is a useful pattern even if you never use Ampersand.

The open question is whether simulated social feedback generalizes to real families and teams. Until independent evaluations and live-traffic results appear, treat the 59.1 and 39.0 percent figures as a hypothesis from the vendor. Related coverage on our site includes the base model's Terminal-Bench 4 result and a look at how multi-agent systems are being built in the open.

Related reading

• GLM-5.3 launch and cyber-defense benchmarks
• GLM-5.3 open-weights delay
• GLM-5.3 ties Kimi K3 on the Intelligence Index
• Hark Pro vs Muse Dots: proactive assistants
• Castform Neon RL post-training
• Self-play with verifiable rewards
• Y Combinator QM multi-agent harness

Details reflect launch coverage as of October 24, 2026. Pricing, availability and benchmark figures may change.