Persimmon: a 550B model that simulates people instead of helping them
humans& put it in a gated preview: it fools a Turing judge 90x more than GPT-6 Astra, near-passes AI detectors, and lets you red-team agents against a crowd.
Copy markdown
Trained to be people, not an assistant
Persimmon is a 550B model — built on NVIDIA's open Nemotron 3 Ultra and fine-tuned with RL against an adaptive discriminator — that reproduces how different individuals actually talk across turns, context, and time, rather than the single obliging voice every other LLM converges on. It ships as a controlled research preview: an open scenario playground now, plus an approval-gated API with academic credit grants. No open weights.
It fools the judges and the detectors
On a Multi-User Turing Test it fooled the LLM judge 19.8% of the time versus 0.22% for GPT-6 Astra (chance is 50%), and the Pangram detector flagged its output as AI just 2.3% of the time — next to a 1.3% human baseline. Coherence decays like a real person's, holding 60.7% at 80 turns.
Build this weekend: stress-test your agent against a crowd
Point a chatbot or multi-user agent at a simulated cohort to generate realistic group-chat eval data, probe moderation and information-sharing, and watch how your product behaves across personalities instead of one cooperative tester. On the Trickle Test, Persimmon hit 88.49% precision and 77.03% recall on who-shares-what. Start in the playground; the API needs human review to unlock.