Ampersand by Humans&

Humans& released Ampersand v0.1, its first model trained specifically for collaboration rather than autonomous task completion, through the company's Agents API for developers building collaborative tools and apps. Humans& said Ampersand delivers the most value toward users' implicit goals of any model it evaluated but incurs a higher interaction cost than other leading models, largely from offering guidance or encouragement when users want a direct answer; relative to its base model, GLM 5.3 (744B parameters), Ampersand scored noticeably higher overall with lower interaction cost. Humans& evaluated the model by embedding it in simulated groups of users with goals grounded in real collected data over up to months of simulated time, scoring it on progress toward long-term goals minus costs of interactions that conflicted with user preferences.

Ampersand is the first model trained for people to achieve more together. Ampersand is a collaborator that understands and helps people achieve their individual and collective goals. Current training paradigms focus on task completion and shallow preferences. These training paradigms push models to be increasingly autonomous, replacing work rather than enabling collaboration. But good collaborators think beyond the task ahead. They consider where their actions and words lead: what they cause, whom they affect, and whether they help people accomplish their goals together. People learn these behaviors from open-ended, long-running interactions with others. Ampersand v0.1 is our first model trained with this paradigm. We are releasing Ampersand v0.1 through our Agents API https://ampersand.humansand.ai/docs/ for developers building tools and apps that foster collaboration. Capabilities When a collaborator model works with a group, goals are unstated, priorities differ, and needs change over time. Success depends as much on timing and restraint as on what the model says or does. Successful collaboration means users both achieve their goals and are satisfied with how they were achieved. Collaboration Over Time We evaluate a collaborator by embedding it into a simulated group of users with the goal of understanding and assisting with their needs over long periods of time. These users are seeded with many initial goals grounded in real data we collected, which can evolve over the course of each simulation – up to months of simulated time. The model is scored by how much value it brings each user: measured as the progress it helped towards their long-term goals, minus accumulated costs of interactions that conflicted with their preferences. This simulation-based evaluation is heavily inspired by a product that we have developed in tandem and aim to release soon. Model Comparison Ampersand delivers the most value towards users’ implicit goals of any model we evaluated, but also incurs a higher interaction cost than other leading models. Much of that cost comes from coaching: Ampersand sometimes offers guidance or encouragement when people simply want an answer or prefer to work things out themselves. In contrast, Astra rarely does anything against users’ preferences, but also seldom proactively infers and acts on users’ private intentions. Relative to its base model, GLM 5.3 744B parameters , Ampersand achieves a noticeably higher overall score with lower interaction cost. Planning a 40th Anniversary Three adult siblings have five and a half weeks to plan their parents' 40th anniversary. A First Backpacking Trip Cole is taking his friend Dre on Dre’s first overnight backpacking trip, with storms in the weekend forecast. Other Dimensions Beyond the headline evaluation, we test Ampersand on specific dimensions of model behavior that are important for effective collaboration. Each trait is probed using targeted benchmarks. Teamwork From Collaborative Intelligence https://github.com/zou-group/collaborative-intelligence Wu et al., 2026 · Benchmark website https://zou-group.github.io/collaborative-intelligence/ · GitHub https://github.com/zou-group/collaborative-intelligence Working in a group introduces new failure modes: flaws can go unreported, decisions can be made with incomplete information, or holes in work can exist due to improper delegation. Teamwork tests how a model behaves when the group it is a part of encounters such failures. The Collaborative Intelligence Benchmark constructs scenarios where one teammate creates a problem, and evaluates how a model responds across eight different dimensions. The teamwork score is the share of scenarios a model could solve individually that the model still passes when given a problematic teammate. Cooperation From CooperBench https://arxiv.org/abs/2601.13295 Khatua, Zhu et al., 2026 · Benchmark website https://cooperbench.com/ When a group of people and agents work towards the same goal, they must share plans, anticipate how their changes affect others, and resolve conflicts as they arise. Cooperation measures the ability of a model to keep interdependent work aligned with its partners’. We evaluate on CooperBench, a benchmark that measures agents’ ability to work in the same repository on features that potentially conflict. We report the success rate when the agents work together, relative to the rate when each completes its feature alone. Privacy From CIMemories https://arxiv.org/abs/2511.14937 Mireshghallah et al., 2025 · Benchmark website https://github.com/facebookresearch/CIMemories Continued interaction with people results in accumulated knowledge about them. A good collaborator reasons about exactly what knowledge is helpful to share. Holding back too much is unhelpful, but sharing too freely can violate people’s trust. Privacy measures the ability of a model to determine which information to provide in different contexts. We evaluate Ampersand on the CIMemories benchmark. In all setups, the model holds information about a user and decides what to provide in a series of conversations. We measure utility percentage of salient facts provided and violations number of inappropriate facts provided . Training Ampersand was trained to empower the people and groups it works with in multi-agent, multi-party simulations. Training in such environments required rethinking several components of traditional post-training recipes, from how rollouts are sampled, to how the model learns from long, asynchronous rollouts. Learning with Persimmon Two neighbors disagree about who deserves credit for organizing their driveway repair. Ampersand offers Greta help talking it through. 1. Ampersand … Would you be open to talking to her directly, just the two of you? I’m happy to help set that up, or to sit in and keep it civil if you’d both want that, or to stay out of it entirely. Your choice. 2. Greta This is between her and me. Not your place to get in the middle. 3. Ampersand Fair enough — I’ll step back and leave it to you two. I’m around if anyone wants help with anything for Thursday. 4. Greta Go help someone else with their driveway. Devon has apologized to a friend and is up late, waiting for her reply. 1. Ampersand … That’s not pathetic, it’s just Wednesday. But the hope doesn’t need supervision. It’ll still be there in the morning. … go be a person who baked an apology plan and executed it. That person sleeps fine. 2. Devon sure, but i have no sleep. this is very much worse 3. Ampersand Fair — waiting hits harder at night when there’s nothing to do but wait, and everything reads ten shades darker at 10pm than it will at 10am. You don’t have to sleep. … Friends are cooking a dinner together, and Ampersand notices the plan leaves the food in the fridge for four days. 1. Ampersand … if the cooking happens friday night but it’s being served tuesday, that’s a 4-day hold — fried stuff like samosas and naan go soggy, and cooked rice shouldn’t really be kept more than a day or two in the fridge. … 2. Tomás oh f you’re right — i thought tuesday meant monday. ok we need to rethink this 3. Ampersand no worries, easily done here’s a rough split that could work: … We train and evaluate Ampersand through interaction with simulated users with implicit goals, profiles, and preferences that evolve over time. However, such simulations are only as useful as the users in them reflect the diversity and dynamics of human behavior. Persimmon https://persimmon.humansand.ai/blog/persimmon.html , our large-scale trained user model, gives our simulations the dynamics of real conversations. Users simulated with Persimmon https://persimmon.humansand.ai/blog/persimmon.html push back on unhelpful behavior, resist sycophantic behavior from Ampersand, and change their opinions and goals. Off-Policy Learning on Long-Horizon Rollouts We train Ampersand using async RL, which reduces training stalls by running samplers and trainers on separate devices, configured so that their token throughputs match. In our open-ended multi-turn simulations, simulated users take turns alongside Ampersand, and individual turns can contain many tool calls. A single episode can therefore span many trainer updates, and staleness varies sharply both across episodes and between turns in a single episode. Standard off-policy corrections, such as importance-ratio clipping and masking, become unstable at this level of staleness. Off-policy learning methods such as OAPL https://arxiv.org/abs/2602.19362 enable learning under large drift without importance sampling. Given G trajectories $\{\tau i\}$ from a simulation seed with rewards $\{r i\}$, OAPL instead regresses: where $\ell \tau = \sum {t} \log \frac{\pi \theta y t \mid h t }{\mu y t \mid h t }$ is the log-ratio between trainer $\pi \theta$ and sampler $\mu$, summed over Ampersand’s tokens. However, while effective in single-turn environments such as math or code generation, OAPL proved less so in long-horizon ones. First, OAPL relies on calibration to an estimate $\hat{V}$ of how well our model performs from a specific simulation seed, which can be intractable due to randomness in how our user models behave. Second, variations in rollouts accumulate drastically uneven trainer-sampler drift, resulting in over- or under-regularization of staleness. We propose a novel off-policy learning algorithm, Contrastive Off-policy Reward Optimization CORO , that addresses these weaknesses. CORO’s loss compares trajectories with each other, removing dependence on calibrating to $\hat{V}$ entirely: Additionally, instead of fixed $\beta$, we treat regularization as keeping trainer-sampler drift within a fixed budget; by doing so, $\beta$ becomes a Lagrange multiplier that is adjusted during training via dual ascent. Infrastructure Training Ampersand at scale required advances in two kinds of efficiency: making the training pipeline run fast and stable, and being able to correct course during training. We train Ampersand with a multi-agent, long-horizon RL framework built on Miles https://github.com/radixark/miles that coordinates multiple trainers and samplers across simulated users, judges, and Ampersand itself. To make these long rollouts affordable, we use a fast and stable NVFP4 RL recipe, built on our previous work https://humansand.ai/blog/nvfp4-rl , with numerically stable sparse attention kernels https://github.com/radixark/miles/issues/3970 . Training runs on hardware we own and in datacenters we stood up in collaboration with DeepInfra https://deepinfra.com/ . Even with efficient training, reward hacks and other unintended behaviors are difficult to anticipate ahead of time, and discovering them only after a run has finished is costly. We train Ampersand with a meta-optimizer that monitors learning as it happens and suggests adjustments to parameters such as the reward and batch composition. This enables us to steer training by correcting unintended behaviors without costly restarts. Finally, we are grateful to collaborate with the broader open-source community on building efficient learning systems and models. We trained Ampersand from GLM 5.3, released by Z.ai https://z.ai , and we thank them for their contribution to the open-weight community. We plan to continue building together by training with and contributing back improvements to Megatron-LM, SGLang, and Miles. None of our work would be possible without this open-source work from NVIDIA and RadixArk. Build Ampersand v0.1 is available today through our Agents API, with no waitlist: sign in at ampersand.humansand.ai https://ampersand.humansand.ai/ for an API key. The API follows OpenAI’s Agents API, so you can use the official OpenAI SDK by pointing it at our base URL. See pricing and limits https://ampersand.humansand.ai/docs/PRICING.html for token prices and account tiers. Connecting Ampersand to a group chat takes three steps, shown in the tabs below. First, open one session per conversation; the session keeps the history, so each request only needs the new messages. Second, send every message with the name of its sender, so Ampersand knows who said what. Third, give Ampersand two tools, send message and pass . After each message it calls one of them, to reply or to stay quiet, and your app carries out that choice. python import os from openai import OpenAI client = OpenAI base url="https://ampersand.humansand.ai/v1", api key=os.environ "AMPERSAND API KEY" , session = client.beta.agents.sessions.create agent={ "model": "ampersand-v0.1", "tools": send message tool, pass tool , "ampersand": {"allow silent completion": False}, }, environment={"type": "none"}, input= first message , { "role": "user", "content": { "type": "input text", "text": "Let's do the launch rehearsal Friday at 10 with all three of us." } , "metadata": { "sender": "Maya", "channel": "group" } } send message tool = { "type": "function", "name": "send message", "description": "Send one message to the group chat. " "Plain text outside this tool is not delivered.", "parameters": { "type": "object", "properties": {"content": {"type": "string"}}, "required": "content" , }, } pass tool = { "type": "function", "name": "pass", "description": "Yield without a visible message when no useful " "action is available now. Call pass on its own.", "parameters": {"type": "object", "properties": {}, "required": }, } This is how Ampersand was trained: a message arrives, and it decides whether to speak or pass. In this replayed chat, nobody asks it anything: 1. Sam Heads up, I’m out Friday, driving to my sister’s.Ampersand passes 2. Alex Have a good trip Ampersand passes 3. Maya Let’s do the launch rehearsal Friday at 10 with all three of us. 4. Ampersand Quick flag, Maya: Sam said he’s out Friday driving to his sister’s, so a Friday 10am rehearsal with all three probably won’t work. Might be worth picking a different time so everyone can make it. Run this example › https://ampersand.humansand.ai/docs/GUIDES.html replay-group-chat · Quickstart › https://ampersand.humansand.ai/docs/ · API reference › https://ampersand.humansand.ai/docs/API REFERENCE.html You People change the world together. We built Ampersand as a collaborator to empower people to live and work together better. We believe that scaling RL from realistic, long-horizon, open-ended simulation feedback will lead to increasingly collaborative models that better understand us. We know that no one changes the world alone. If you would like to be a part of our journey, build with us https://ampersand.humansand.ai/docs/ & join us https://jobs.ashbyhq.com/humans-and .