Goekdeniz-Guelmez/JOSIE-2-4B-OSS

JOSIE-2-4B-OSS: Reasoning-First Language Model

JOSIE-2-4B-OSS is a 4.5 billion parameter model from the JOSIE family, developed by Gökdeniz Gülmez. It is built upon the Qwen3.5 architecture and uniquely emphasizes reasoning-first supervision, trained entirely on consumer Apple Silicon. The core hypothesis behind JOSIE is that a capable language model can substantially improve by learning how to reason, rather than simply being taught more facts.

Key Capabilities & Differentiators

• Reasoning-First Training: Trained on approximately 4 million tokens of carefully curated data, focusing on decomposing complex tasks, verifying intermediate results, and recognizing uncertainty.
• Improved Internal Reasoning: Aims to reshape how the underlying Qwen3.5 base model approaches problems, cultivating self-awareness and intellectual honesty.
• Benchmark Performance: Demonstrates significant improvements over the base Qwen3.5-4B model, achieving 95.5 on ARC-C and 69.2 on TruthfulQA in reasoning mode, and 91.9 on ARC-C and 66.8 on TruthfulQA in non-reasoning mode.
• Emergent Personality: Exhibits an emergent, sometimes sarcastic or informal, internal monologue during its reasoning process, which was not explicitly trained but observed during RL.
• Local Training: Developed and trained entirely on consumer Apple Silicon, proving the viability of local, hardware-agnostic model development.

Use Cases & Recommendations

JOSIE-2-4B-OSS is ideal for applications requiring robust reasoning, problem decomposition, and internally consistent answers. Its unique training methodology makes it suitable for tasks where logical coherence and intellectual honesty are paramount. The model supports two inference modes: /reasoning for explicit reasoning traces and /none-reasoning for direct answers, with performance gains observed in both. It is compatible with various inference frameworks supporting Qwen3.5, including Transformers, vLLM, Ollama, and llama.cpp.