8B (dense)
32,768 tokens native; production deployment uses 16,384
Not specified
Open-weight experimental character model created by fine-tuning Qwen3-8B-Base to embody one narrow, persistent, anti-servile character rather than a general-purpose assistant. Zero represents a man voluntarily enclosed in a white plastic box with no personal memories before November 5, 2021 and only five objects in his environment. Moving Castles defines the character through a human-authored association graph containing 117 nodes and 298 edges, supplemented by separate core-personality and writing-style anchors. Its synthetic supervised corpus was generated through multi-turn conversations between independently sampled model families, without system messages, and contains 5,932 post-filter conversations, 73,765 character turns, and approximately 3.04 million character tokens. Conversations span 10 to 40 turns so that long-horizon character persistence remains in distribution. Supervised training began from the non-instruction-tuned Qwen3-8B-Base checkpoint and used three epochs of rank-64, alpha-128 LoRA training together with fully trained token embeddings and LM head, for 1.42B trainable parameters; the adapter was subsequently merged. Reinforcement learning then used 380 prompts harvested from observed character failures and successes, comprising 193 capitulation onsets, 107 pressure-held examples, and 80 ordinary in-world openers. This stage ran 300 steps of GRPO with DAPO loss modifications, four rollouts per prompt, a rank-16, alpha-32 LoRA adapter, and 43.6M trainable parameters. Its reward combines a Qwen3.6-27B LLM judge scoring voice, world frame, personality, and engagement against the character bible with a penalty for excessive four-gram repetition. In a held-out evaluation of 250 matched conversations with 16 turns each, the merged SFT-plus-RL model exhibited hard character breaks in 2.8 percent of conversations, compared with 22.8 percent for the SFT-only checkpoint and 45.2 percent for the sibling Qwen3-8B instruct model controlled through a system prompt. The final model had zero complete unrecovered collapses, compared with 21 for SFT-only and 60 for prompting. This specialization substantially damages conventional capabilities: Zero scored 2.5 percent on GSM8K versus 62.2 percent for Qwen3-8B-Base, and 20.7 percent prompt-level and 32.7 percent instruction-level strict IFEval versus 46.8 and 58.2 percent for the base model. Mechanistic experiments report that the residual-stream displacement produced by SFT and RL is highly aligned at mean cosine 0.985, while the trained character directions are nearly orthogonal to the system-prompt direction at -0.022 and -0.004. Steering the base model with the fine-tuning displacement elicited traces of Zero, whereas steering with the prompt displacement did not, although the authors caution that orthogonality alone is weak evidence in high-dimensional space. The model should be used without a system prompt because training includes only user and assistant turns; it also requires both ChatML end-of-turn and end-of-text tokens as stopping conditions at nonzero temperature. The checkpoint is distributed in BF16 across four safetensors shards and uses the standard Qwen3 architecture. No explicit license is currently declared for the released checkpoint.