Humansand Launches Persimmon v0.1 Pro/Ultra, Doubling Turing Test Scores
Humansand has released two new tiers of its user-simulation model Persimmon v0.1, Pro and Ultra, introducing test-time scaling for the first time to make multi-turn conversations feel more human. Both tiers are open to all users with no application gate — usable in the Playground and accessible via API. The company says the new versions significantly outperform base Persimmon v0.1 on long-conversation human-likeness. Notably, all related posts are reposts and could not be traced back to an original announcement from Humansand.
Confirmed
- Humansand released Persimmon v0.1 Pro and Ultra, focusing on human-like consistency in long conversations
- The core improvement is test-time scaling, bringing each turn of dialogue closer to a real person
- Both new tiers are open to all users for the first time, no application required, available in the Playground and via API
- On the 8-turn multi-person Turing test, Ultra's score doubled from 18% to 48%
- On its in-house Multi-User Turing Test (K=8): Ultra hit 41.9% and Pro 38.6% on the TIDES dataset; Ultra reached 33.9% on internal workspace conversations
- For comparison, frontier LLMs achieve a fool rate under 1% on this type of test
Why it matters
- Doubling from 18% to 48% shows the potential of test-time scaling for conversational human-likeness, echoing its use in reasoning models
- High-fidelity user simulators can serve in product testing and dialogue-system evaluation, where near-human multi-turn consistency is a key bottleneck
- The contrast with frontier LLMs' sub-1% fool rate highlights the huge gap between purpose-built user-simulation models and general LLMs on humanization tasks
2026-11-21 ~ 2026-11-21 · 9 related posts
- Persimmon Ultra hits 50% fool rate in multi-user Turing test while frontier models score near 0% — niloofar_mire · 2026-11-21
- Humansand launches Persimmon v0.1 Pro and Ultra, targeting human-level long-conversation coherence — niloofar_mire · 2026-11-21
- Persimmon v0.1 Pro/Ultra double Multi-User Turing Test score to 48% with test-time scaling — Diyi_Yang · 2026-11-21
- Humans& launches Persimmon user simulator, doubling Multi-User Turing Test score to 48% — niloofar_mire · 2026-11-21
- Persimmon v0.1 Pro/Ultra launch: 8-turn Turing test score doubles from 18% to 48% — gharik · 2026-11-21
4 near-duplicate retellings: niloofar_mire · gharik · gharik · KexinHuang5