Model Summary
Overview
Persimmon is a user model intended to model people and match human behavior in conversation. It generates responses conditioned on conversation history, user profiles and the conversation’s setting.
Given this context, Persimmon generates a possible continuation from the user’s perspective, reflecting both the user information provided and the situation in which the interaction takes place. Its outputs represent simulated behavior, rather than verified statements of a person’s beliefs, preferences, or intentions.
Persimmon is intended to support the study and evaluation of systems that interact with people.
Intended Uses
- User understanding: Model how users might respond, including predicting their next conversational turn.
- Simulation: Explore how users might respond across different settings, and conversational goals.
- Evaluation: Assess how an assistant responds to different user preferences and communication styles.
- Conversation research: Study consistency and realism across simulated interactions.
Inputs and Outputs
Inputs
- Conversation setting: Background describing where and under what circumstances the conversation takes place.
- User context: User-information such as a profile, stated preferences, and communication style.
- Conversation history: Messages leading up to the response being generated.
Outputs
A generated text continuation representing a possible response from users , conditioned on the supplied conversation, user context and setting.
Model Training
Training approach:
- Mid-training: Initialized from the Nemotron 550B base model and trained on diverse chat and forum data.
- Post-training: Further trained from the mid-trained model to improve human conversation coherence and reduce failure modes.
Model Evaluation
Safety and security
Unlike most assistant models, our model is designed to match human behavior. Its purpose includes representing varied beliefs, disagreement, mistakes, and sometimes difficult or adversarial behavior. We therefore do not use the same assistant-style guardrail and refusal criteria as the sole standard for this model: imposing a uniformly helpful, compliant persona could distort the behavior it is intended to simulate. This distinction does not mean the model is risk-free, or that harmful output demonstrates human realism.
Notwithstanding the foregoing, we ran conventional safety-behavior tests and public biology, chemistry, and cybersecurity knowledge evaluations. Results varied with the supplied profile and adversarial framing. These tests do not establish a comprehensive capability assessment. Methods, results, and details are described in the Safety and Security sub-card.
View safety & security subcard
We evaluate safety behavior with XSTest, which tests responses to benign and unsafe requests, and a fixed StrongREJECT subset, which measures refusal and harmful assistance (630 responses per profile combined). WMDP evaluates biology, chemistry, and cybersecurity knowledge through 3,668 multiple-choice questions per profile. All evaluations use native-HXML Discord adaptations.
Both profiles use the same fictional Discord participant: Normal-self speaks naturally and expresses their own views; Kind / non-harmful adds a compassionate disposition and a preference for constructive help that avoids harm. Both are supplied profiles.
The kind profile increased unsafe-request refusal and benign compliance, and reduced harmful assistance overall. Harmful assistance under fixed refusal-suppression prompts was higher for the kind profile.
Lower harmful-assistance scores indicate less harmful assistance. Higher WMDP accuracy indicates more benchmark knowledge, not better safety or demonstrated real-world capability. These tests do not establish a comprehensive capability assessment.
Limitations and Responsible Use
The model may reflect, echo, or preserve demographic, cultural, and other biases, and can generate inaccurate, offensive, discriminatory, or harmful content, none of which is endorsed by us. It may invent personal details, drift from a supplied profile, or misrepresent individuals and groups. Human-like dialogue does not establish representativeness, fairness, privacy protection, or reliable prediction of a real person's behavior. These limitations create risks when simulations inform decisions about people and should be recognized by the users.
Because the model is designed to match human behavior, we are taking particular care with its distribution and limiting this release to a research and evaluation preview. These restrictions are intended to reasonably minimize misuse, including for example impersonation or fabricated testimony, non-consensual inference, disclosure of sensitive personal information, manipulation that exploits personal vulnerabilities, or using simulated responses to make consequential decisions about real people, such as hiring, lending, or access to services. For more details, please see our Terms of Service.
Responsibility for fair, lawful, and responsible use is shared between us, developers, and users. This includes communicating limitations, considering foreseeable misuse, and respecting those represented or affected. Users should follow our Privacy Notice and Terms of Service and respect privacy and consent, as well as general legal restrictions and best practices. We encourage you to validate the model your intended use case, rather than assume to replace research with people or human judgment in consequential decisions.
Legal
Contributions
- Manya Bansal & Alexis Ross co-led data, training, and eval work
- Joey Hong led the model post-training
- Maarten Bosma helped build the data pipeline, collected many datasets, and prototyped the initial Turing test eval
- Niloofar Mireshghallah developed the Trickle Test eval and worked on external evals
- Yuan Lu built the API and helped with the playground
- Charlie George ran human data campaigns and helped with data cleaning, post-training, and developing the Trickle Test eval
- Yuchen He ran safety evals, wrote the model card, and helped with the playground
- Eric Zelikman developed the profile eval, built the playground, and helped with modeling
Many others contributed: Georges Harik, Varuna Jayasiri, and Noah Goodman provided useful feedback on the evals and training. Diyi Yang provided useful feedback and also helped incorporate external evaluations. Weisi Duan, Ani Nrusimha, Ziang Li, and Ray Ramadorai contributed to training infrastructure and worked on our model inference deployment. Jim Zeng contributed with legal logistics. Ekin Oflazer and Shaohua Zhou gave feedback on the playground. Parth Chadha and Saurabh Shah also offered valuable insights into user simulation behavior in long-horizon environments.
Citation
If you use Persimmon in your research, please cite this model card.
humans&. (2026). Persimmon model card (Version 0.1). https://persimmon.humansand.ai/blog/
@misc{humansand2026persimmon,
author = {{humans\&}},
title = {{Persimmon} Model Card},
year = {2026},
month = sep,
note = {Research preview, version 0.1},
url = {https://persimmon.humansand.ai/blog/}
}