Email-Qwen3-0.6B: Specialized Email Generation

This model, built upon the Qwen3 0.6B architecture, is meticulously fine-tuned by CharlieGreenman for generating professional email bodies. It leverages a unique training methodology involving 130,000 curated prompt-email pairs and 5 rounds of rejection sampling, ensuring high-quality output.

Key Capabilities

• Professional Email Generation: Creates polished email bodies from concise prompts.
• Broad Email Type Support: Handles over 17 types, including cold outreach, follow-up, thank-you, request, apology, and invitation.
• Optimized for Small Scale: A 0.8 billion parameter model, making it efficient for local deployment, with a quantized Q4_K_M version (378MB) available via llama.cpp.
• Context Length: Supports a context length of 32768 tokens.

Good For

• Automated Email Drafting: Ideal for quickly generating various email types for business or personal use.
• Resource-Constrained Environments: Its small size allows for efficient inference, especially with quantized versions.
• Integration into Applications: Designed for easy integration using llama.cpp or Hugging Face Transformers.

Limitations

• Best suited for common email types; may struggle with highly unusual or creative requests.
• Primarily generates email body text; subject lines need separate handling.
• As a smaller model, quality benefits significantly from techniques like best-of-N selection and post-processing.
• May occasionally produce minor factual inaccuracies like hallucinating company names or statistics.