10 local LLMs, one brand prompt, 8 setups: what I learned generating product copy on an ASUS GX10
Description and use
Automated generation of German product copy for a B2B shop with 10,000+ items on a local ASUS GX10. Gemma 4 26B scored highest with a plain prompt; brand prompt lowered scores. 4‑bit QAT preserved quality and speed, while vLLM batching cut runtime to 1.6 days.
Problem
Generate correct German product descriptions for a B2B workwear shop with 10,000+ articles, adhering to brand rules and avoiding invented properties, on a local machine without discrete GPU.
Outcome
Gemma 4 26B achieved the highest score (288/300) with plain prompt; brand prompt reduced scores by ~60 points. Quantization-aware training (4‑bit QAT) maintained quality while improving speed; post‑hoc 4‑bit quant lost 22 points. Batching with vLLM ×8 reduced generation time from 5.4 to 1.6 days for 10,000 articles.
Workflow
- Collect supplier data and product metadata
- Clean data in code before prompting
- Apply fixed rules for pattern-based facts
- Prompt LLM with brand voice prompt
- LLM rewrites only flagged sentences
- Validate output against rubric with Claude reviewer
- Iterate on model build and quantization
- Batch requests with vLLM for speed
- Store generated HTML and SEO fields
- Human spot‑check for quality assurance
Evidence
Ten local models, 30 articles, same data, same prompt. The Gemma 4 family won (Gemma 4 26B: 288/300, with 12B and 31B right behind).
Scoring was stricter because tone counted, so look at the spread: 7 points with the plain prompt, 111 with the real one.
Full precision bought nothing. BF16 was slower and not better.
Batching is the speed lever. vLLM with 8 concurrent requests instead of Ollama with 2: 10,000 articles in about 1.6 days instead of 5.4.
- Record date: 2026-10-22T00:26:20.759Z