Models

humans& debuts Persimmon model to mimic human chat

AI startup humans& has released Persimmon, a 550-billion parameter model designed to realistically simulate human behavior in group chats to improve agent evaluations and synthetic data.

AlphaSignal16 hrs agoModels
Image: AlphaSignal

AI startup humans& has introduced Persimmon, a 550-billion parameter model built to simulate realistic human behavior in multi-turn group chats. Unlike traditional assistant models optimized to be helpful, Persimmon mimics human quirks, hesitation, and conversational drift. The model is initialized from NVIDIA's Nemotron 3 Ultra 550B base, midtrained on public internet dialogues, and post-trained using reinforcement learning against an adaptive discriminator on thousands of Blackwell GPUs.

In a Multi-User Turing Test evaluated by an LLM judge, Persimmon fooled the judge 18% to 21% of the time, whereas frontier assistants scored under 3%. On the TIDES dataset, Persimmon achieved a 21.1% fooled rate, compared to 2.7% for Nemotron Ultra Base, 0.11% for GPT 5.6, 0% for GPT 6 (Astra), and 0.11% for Opus. For internal workspace conversations, Persimmon scored 18.6% while all competitors fell below 1%. On TutorMoments, it reached 19.8%, far ahead of GPT 6 (Astra) at 0.39%. Furthermore, the Pangram AI detector flagged Persimmon transcripts as AI-generated only 2.3% of the time, close to the human baseline of 1.3%, while flagging GPT 6 (Astra) at 54.3% and Fable 5 at 94.3%.

Persimmon also excels at pacing information. On the Trickle Test, which measures how naturally a model reveals facts, Persimmon achieved the highest precision among 43 tested models at 88.5%, with a recall of 77%. However, it still struggles with long-context coherence. Over 80-turn conversations, humans remain coherent 87.3% of the time, while Persimmon drops to 60.7%. In contrast, GPT 6 (Astra) and Opus 5 scored 98.4% and 99% respectively, which researchers note is actually too coherent compared to real, forgetful humans.

For developers, Persimmon offers a way to close the realism gap in agent evaluations, negotiation training, and synthetic dataset generation. Instead of relying on prompt engineering to simulate users, practitioners can now access a dedicated model that behaves like a real person. humans& is offering a gated research preview of Persimmon through a playground and an API, with academic credit grants available to approved applicants.

This is our own summary of reporting by AlphaSignal

More in Models