Madras1/Jade-14B

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:14BQuant:FP8Ctx Length:32kPublished:Apr 7, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

Madras1/Jade-14B is a Brazilian Portuguese conversational finetune of the Qwen3-14B model, developed by Madras1. This model is specifically designed to express a strong, persistent persona, excelling in PT-BR chat, chatbot use cases, and character-style interactions. It consistently uses colloquial language, abbreviations, slang, and a WhatsApp-like tone, making it ideal for informal and friendly conversational agents.

Loading preview...

Overview

Madras1/Jade-14B is a 14 billion parameter conversational language model, a finetune of unsloth/qwen3-14b, developed by Madras1. Its primary distinction is its persona-first design, where the model speaks with a consistent "Jade" persona even without explicit system prompts. This model is optimized for Brazilian Portuguese (pt-BR) and is licensed under apache-2.0.

Key Capabilities

  • Consistent Persona: Maintains a strong, informal, and friendly persona in all interactions.
  • Natural PT-BR: Speaks naturally in Brazilian Portuguese, incorporating colloquialisms, slang (e.g., vc, tmj, mano), and abbreviations.
  • Chat-Oriented: Designed for casual chat, chatbot use cases, and WhatsApp-style conversations.
  • Persona in Weights: The finetuning objective was to embed the persona directly into the model's weights, ensuring personality from the first turn.

Training Details

The model was trained using a supervised finetuning (SFT) approach on approximately 25,000 examples over 3 epochs. The training involved synthetic PT-BR prompt generation for chat-like situations and persona-driven response distillation, with system persona instructions removed during SFT to internalize the Jade style.

Recommended Use

  • PT-BR chat assistants
  • Persona bots and character-style interaction
  • WhatsApp-style conversational agents
  • Lightweight entertainment or social AI experiences

It is less suitable for formal writing, neutral assistant behavior, or high-stakes contexts like legal or medical applications due to its inherent informality.