Goekdeniz-Guelmez/JOSIE-2-9B-OSS
JOSIE-2-9B-OSS is a 9 billion parameter language model from the JOSIE-2 family, built on the Qwen3.5 architecture by Gökdeniz Gülmez. It is uniquely trained on consumer Apple Silicon with a reasoning-first approach, focusing on decomposing complex tasks, verifying intermediate results, and producing internally consistent answers. This model excels at improving internal reasoning structure and cultivating self-awareness, demonstrating significant performance improvements over its base model in both reasoning and direct-answer modes.
Loading preview...
Overview
JOSIE-2-9B-OSS is part of the JOSIE-2 model family, developed by Gökdeniz Gülmez. This model is built upon the Qwen3.5 architecture and is notable for being trained entirely on consumer Apple Silicon. The core hypothesis behind JOSIE-2 is to explore whether a language model can substantially improve by learning how to reason, rather than simply being taught more facts. It focuses on reasoning-first supervision, aiming to reshape how the underlying Qwen3.5 base models approach problems by emphasizing task decomposition, verification of intermediate results, and recognition of uncertainty.
Key Capabilities & Differentiators
- Reasoning-First Training: Trained on approximately 4 million tokens of carefully curated data, with a strong emphasis on improving internal reasoning structure and cultivating self-awareness.
- Performance Improvement: Demonstrates significant improvements over the base Qwen3.5 model in both reasoning and non-reasoning modes, particularly in benchmarks like ARC-Challenge and TruthfulQA.
- Emergent Personality: Exhibits an emergent, sometimes sarcastic or informal, internal monologue during reasoning, which is an ongoing research observation.
- Local Training: Developed and trained entirely on consumer Apple Silicon, proving the viability of local, hardware-agnostic model development.
When to Use This Model
JOSIE-2-9B-OSS is particularly well-suited for applications requiring:
- Enhanced Reasoning: Tasks that benefit from structured problem decomposition, verification, and intellectually honest responses.
- Consistent Personality: Use cases where a model with a distinct, self-aware, and task-loyal personality is desired.
- Internal Consistency: Scenarios where producing internally consistent answers, even when facing uncertainty, is critical.
- Qwen3.5 Compatibility: Any environment or framework that supports Qwen3.5 models, including Hugging Face Transformers, vLLM, Ollama, and llama.cpp.