Anyone who writes advertising copy knows that general-purpose models, even the most capable ones, tend to fall back on formulaic phrases. Openings like “In today’s fast-paced world,” vague promises, and weak calls to action are the hallmark of copy produced without proper training in the craft.

A project shared on Reddit shows what happens when you tackle the problem with a surgical fine-tune. Using Gemma-4-31B-it as a base, the team curated a corpus of creative briefs paired with final copy, including Facebook ads, cold emails, landing pages, and video scripts. Training was done with QLoRA, so computational cost remained low, and the final weights were merged to full bf16 to simplify deployment. The resulting model has 31 billion parameters and a 256K token context window—two features that make it usable on high-end consumer hardware without too many compromises.

To measure progress, a custom benchmark was built on the EQ-Bench 3 methodology: 30 real briefs, judged blind in pairwise comparisons by DeepSeek V4 Flash, with position bias controls. The verdict is clear: the fine-tuned model scores 1657 Elo versus 1367 for the base version, winning 80% of head-to-head matches. The biggest gains are exactly where direct-response copywriting shines: hook strength, specificity, and concision.

This is more than a creative tuning exercise. For those running text generation workloads locally, it signals the maturity of open models when specialized for vertical domains. Gemma 4 uses less VRAM than larger alternatives, and the quantized version available on Hugging Face (tagged base_model:quantized:akwin123/copywriter-gemma4-31b) opens up inference on GPUs with 24–48 GB of memory—an interesting margin for marketing departments that want full data control, avoiding cloud APIs and their privacy unknowns.

One operational detail matters: the model must be used with enable_thinking=false. Turning on Gemma 4’s reasoning mode, paradoxically, hurts output quality. This isn’t surprising to those who work with instruction-tuned models: the so-called thinking mode can introduce reasoning chains that pull away from clean execution of a creative brief, a trade-off reminiscent of debates about overly complex prompts in RAG pipelines.

Integration is designed to be straightforward: the weights are released in Transformers format and compatible with vLLM, with no separate adapter to manage. You simply load the model and point it at a text generation pipeline. Anyone looking for ideas for a local stack will find loading examples and decoding tips in the model card on Hugging Face.

The point isn’t whether AI can write better headlines than an experienced copywriter. It’s that a model with this kind of specific training becomes a reliable workmate for large-scale A/B testing or industrial-volume variant generation—without sending data to outside servers.