Gryphe/WorldSim-Opus-3.6-35B-A3B
Gryphe/WorldSim-Opus-3.6-35B-A3B is a 35.1 billion parameter Mixture-of-Experts (MoE) model based on Qwen 3.6, developed by Gryphe. This model is an experimental fusion of creative world simulation and genuine reasoning capabilities, fine-tuned on datasets that include full thinking traces for planning story beats and character motivations. It excels at long-form narrative roleplay and instruction-following by reasoning through creative writing tasks.
Loading preview...
Overview
Gryphe/WorldSim-Opus-3.6-35B-A3B is a 35.1 billion parameter Mixture-of-Experts (MoE) model, built upon the Qwen 3.6 architecture. Developed by Gryphe, this model represents an experimental approach to combine creative world simulation with robust reasoning. Its core differentiator lies in its training methodology, which incorporates datasets featuring extensive thinking traces, allowing the model to plan and reason through creative writing and narrative generation.
Key Capabilities
- Reasoning-driven Creative Writing: The model is trained to generate creative content by planning story beats, considering character motivations, and working through consequences, as evidenced by its internal thinking traces.
- Long-form Narrative Roleplay: It excels in extended storytelling and character immersion, leveraging "WorldSim data" focused on emergent world logic.
- Instruction Following: Incorporates general instruction-following capabilities from cleaned Claude Opus 4.6 reasoning traces.
- Reduced AI Clichés: Training data includes critic-improver rewrites to mitigate common AI-generated clichés.
Training and Unique Features
The model was fine-tuned on a blend of reasoning datasets, including Opus-4.6-Reasoning-24k (50%), WorldSim data (40%), and Tiamat data (10%). A notable feature is the use of preserve_thinking: true during training, ensuring thinking tags are active across all assistant turns in multi-turn conversations. This approach aims to improve the model's ability to reason consistently throughout a dialogue. The base model, Qwen3.6-35B-A3B, was chosen for its MoE architecture and ease of training with Axolotl.