JOSIE-4B-Instruct: Uncensored, Human-like AI Assistant

JOSIE-4B-Instruct is a 4 billion parameter, full-weight fine-tuned instruction-following model developed by Gökdeniz Gülmez. It is built on a gabliterated version of Qwen3-4B-Instruct-2507, meaning it provides uncensored and direct responses without artificial constraints. The model is designed for natural conversational interactions, problem-solving, and general assistance, featuring a friendly, direct, and naturally expressive personality with appropriate emoji use.

Key Capabilities & Characteristics

• Uncensored Output: Provides direct and unfiltered responses, free from built-in content filtering.
• Human-like Personality: Engages in natural, expressive conversations, offering practical help without being sycophantic.
• Strong Instruction Following: Adheres well to user instructions and preferences across diverse tasks.
• Multilingual Support: Capable in English, Spanish, French, Portuguese, Italian, Arabic, Japanese, Korean, Indonesian, Russian, Vietnamese, German, and Thai.
• Optimized for Apple Silicon: Trained using MLX frameworks, demonstrating viability on consumer hardware.
• Context Length: Supports a 32,768-token context window.

Ideal Use Cases

• Conversational AI: For chatbots and virtual assistants requiring engaging dialogue.
• General Assistance: Practical help with everyday tasks, coding, analysis, and writing.
• Content Generation: Creative writing, brainstorming, and ideation with a distinct personality.
• Educational Support: Tutoring and explanations in an accessible, friendly manner.

Comparison with JOSIE-4B-Thinking

While JOSIE-4B-Instruct excels in natural conversations and general assistance, its counterpart, JOSIE-4B-Thinking, is optimized for complex reasoning tasks with a longer 65K token context and structured reasoning chains. Choose JOSIE-4B-Instruct for everyday interactions and engaging chat experiences.