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

VISIONConcurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 31, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

JOSIE-2-2B-OSS is a 2 billion parameter causal language model developed by Gökdeniz Gülmez, built on the Qwen3.5 architecture. This model family is uniquely trained entirely on consumer Apple Silicon, focusing on reasoning-first supervision to enhance structured reasoning, consistent personality, and self-awareness. It excels at decomposing complex tasks, verifying intermediate results, and producing internally consistent answers, demonstrating improved performance in both reasoning and direct-answer modes over its base model.

Loading preview...

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

JOSIE-2-2B-OSS is a 2 billion parameter model from the JOSIE family, developed by Gökdeniz Gülmez. A key differentiator is its training methodology: it's entirely developed on consumer Apple Silicon and emphasizes reasoning-first supervision. This approach aims to improve how the model approaches problems rather than just expanding its factual knowledge.

Key Capabilities & Features

  • Enhanced Reasoning: Trained on approximately 4 million tokens of carefully curated data, focusing on decomposing complex tasks, verifying intermediate results, and recognizing uncertainty.
  • Consistent Personality & Self-Awareness: Designed to exhibit a consistent personality and self-awareness, with a system prompt that encourages a "sharp, no-BS, confident" persona.
  • Performance Improvements: Benchmarks show significant improvements over the base Qwen3.5-2B model in both /reasoning and /none-reasoning modes on ARC-Challenge and TruthfulQA, despite being trained exclusively on reasoning-trace demonstrations.
  • Broad Compatibility: Built on the Qwen3.5 architecture, making it compatible with various inference frameworks like Transformers, vLLM, MLX, and Ollama.

Unique Research Focus

The JOSIE project explores whether reasoning-first supervision fundamentally improves a model's internal reasoning policy, or if gains are primarily due to high-quality training data. The observed improvements in direct-answer mode suggest that this training approach may influence how the model internally structures and applies existing knowledge.

Recommended Use Cases

This model is particularly well-suited for applications requiring:

  • Structured Problem Solving: Where the ability to break down problems and verify steps is crucial.
  • Consistent Conversational Agents: For scenarios needing a distinct and reliable model persona.
  • Research into LLM Reasoning: As a valuable tool for exploring the impact of reasoning-first training on model behavior.

Finetunes
6 models