Gryphe's Pantheon-Reasoning-27B brings thinking traces to uncensored roleplay
An experimental 27B model that uses full reasoning traces for character-driven narrative.
Gryphe's experimental Pantheon-Reasoning-27B is a successor to the Pantheon roleplay series and the Codex release, designed to bring reasoning capability to character-driven narrative. Built on llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved, the model uses full thinking traces across every assistant turn, enabling it to weigh tone, plan narrative beats, and consider character authenticity before generating a line. The goal is to determine whether reasoning traces can meaningfully improve roleplay quality over traditional non-reasoning models. GGUF quantizations are available for local use.
Training data is composed of multiple curated sources, each with reasoning traces: ~28% from the core Pantheon roleplay corpus, ~21% from cleaned Opus 4.6 reasoning traces (covering STEM, coding, and instruction-following), ~16% WorldSim narrative roleplay (extended storytelling, emergent world logic), ~16% text adventure content (high stakes interactive fiction), ~16% general roleplay transcripts, and ~3% from the Tiamat dataset (multi-step rewrites to reduce AI clichés). The model's preserve_thinking: true setting keeps reasoning tags active in multi-turn conversations, not just the first turn, making it suitable for prolonged character interactions.
- Pantheon-Reasoning-27B uses a Qwen 3.6 27B uncensored base with full thinking traces for every assistant turn.
- Training data composition: 28% Pantheon roleplay, 21% Opus 4.6 reasoning traces, 16% WorldSim narrative, 16% text adventure, 16% general roleplay, 3% Tiamat data.
- Model preserves thinking tags across multi-turn conversations via preserve_thinking: true; GGUF quants are available for local deployment.
Why It Matters
Brings structured reasoning to uncensored roleplay, potentially improving narrative coherence and character depth for AI storytellers.