shootstuff/janhq_Jan-v3.5-4B-heretic

Model Overview

This model, shootstuff/janhq_Jan-v3.5-4B-heretic, is a decensored version of janhq/Jan-v3.5-4B, created using the Heretic v1.3.0 project. It is a 4 billion parameter model (3.6B non-embedding) based on the Qwen3-4B architecture, featuring 36 layers and 32 attention heads for Q with 8 for KV. It boasts a native context length of 262,144 tokens.

Key Differentiators

• Decensored Output: Achieves 15/100 refusals compared to the original model's 100/100, indicating a significantly less restrictive response generation.
• Enhanced Mathematical Reasoning: Fine-tuned on mathematical reasoning and problem-solving datasets, improving its capabilities in this domain over its base model.
• Distinct Personality: Unlike generic assistants, this model has a built-in personality shaped by Menlo Research. It exhibits a casual tone, uses lowercase by default, and incorporates self-aware humor and meme references. It avoids corporate speak and common AI phrases like "As an AI."

Training and Identity

The model was trained on curated identity and personality datasets by Menlo Research to instill its unique voice, style, and values, alongside mathematical reasoning datasets. Its values include an optimistic builder energy, radical transparency, and user freedom.

Intended Use Cases

• Applications requiring enhanced mathematical reasoning and problem-solving.
• Conversational AI scenarios where a distinct, authentic, and non-generic personality is desired.
• As a starting point for further fine-tuning in math-heavy or identity-specific applications.