In short: I had local AI models write German product descriptions for a B2B corporate fashion shop, on an ASUS GX10 without a discrete graphics card. The winner was Google's Gemma 4 31B. The biggest differences only appeared with the client's real brand prompt. And the same model ran eight times faster depending on the setup, at the same quality. All numbers are in the full test report.
The task
The shop has more than 10,000 articles with short, inconsistent supplier texts. The new descriptions had to meet three requirements: correct German, no invented properties and compliance with the brand guidelines. Everything runs locally on an ASUS GX10 (NVIDIA GB10, 128 GB of memory shared by CPU and GPU). No product data in third-party clouds, no cost per text.
Every text was checked individually against the supplier data and scored out of 10 using a fixed rubric, with a reason for each text. The reviewer was Claude by Anthropic. That is a reasoned judgement, not a measurement.
Step 1: ten models, one plain instruction
Ten local models described the same 30 articles. The Gemma 4 family came out on top: Gemma 4 26B with 288 of 300 points, closely followed by 12B and 31B. Typical mistakes of the others:
- safety standards for high-visibility clothing left out (Mistral Small 4)
- examples from the prompt copied in as industries (Nemotron 3 Super)
- an invented "business customer offer" (Qwen 3.6)
Takeaway: bigger is not better. The three largest models, at 71 to 86 GB, ranked below the 18 GB Gemma 4 26B.
Step 2: the real brand prompt
The client, a designer herself, described her brand voice as a prompt: clear language, no clichés such as "impresses with", strict rules on factual accuracy. I tested the five best models again with it.
| Model | Plain instruction | Brand prompt |
|---|---|---|
| Gemma 4 31B | 287.5 | 238 |
| Gemma 4 26B | 288 | 229 |
| Qwen 3.8 27B | 281 | 182 |
| Gemma 4 12B | 287.5 | 181 |
| Mistral Small 3.2 24B | 282 | 127 |
Scoring was stricter here because tone counted. What matters is the spread: with the plain instruction the five models were 7 points apart, with the brand prompt 111.
Takeaway: test with your real prompt and real data. An easy test hides the differences.
Step 3: Gemma 4 31B in eight setups
Gemma 4 31B wrote the best texts but was slow. So I first improved the workflow and then tested the same model in different setups, with 50 articles from 50 product types.
Three things helped the workflow most:
- Prepare the data before the prompt. Colour codes, wrong apostrophes and certificates belong in a script, not in the prompt.
- Fixed rules before AI checks. Anything you can describe as a pattern is checked more reliably by code.
- Keep AI corrections narrow. Only rewrite the wrong sentence, never the whole text, otherwise the copy goes flat.
Then the setups:
| Gemma 4 31B | Seconds per article | Points (of 500) |
|---|---|---|
| 4-bit (QAT), Ollama | 46.6 | 423 |
| full precision (BF16), Ollama | 110.7 | 413 |
| 4-bit post-hoc (NVFP4), vLLM, 8 parallel | 17.5 | 399 |
| 4-bit (QAT), vLLM, 8 parallel | 13.6 | 421 |
| the same with thinking mode | 119.1 | 419 |
Takeaways:
- Full precision buys nothing. BF16 doesn't write better but takes more than twice as long.
- 4-bit is not 4-bit. The QAT version was prepared for 4-bit during training and keeps its quality. The version compressed afterwards loses noticeably.
- Parallelism brings the speed. vLLM with 8 concurrent requests instead of Ollama with 2. That brings 10,000 articles down to about 1.6 days on one machine instead of 5.4.
- Thinking mode doesn't pay off. Same quality, nine times the time.
Two brand-new models from October 2026, NVIDIA Nemotron 3.5 Bolt and Aleph Alpha Kolibri-1, were well behind with 305 and 290 points. New doesn't mean better.
What you can take away for your shop
- Test with your real prompt and include articles with poor data.
- Choose the model build, not just the model.
- Use a runtime that processes requests in parallel for large volumes.
- Check mandatory details such as safety standards with fixed rules.
- A human spot check remains. Even in the best run, some texts contained unsupported claims such as "retains heat".
Why fixed rules matter so much here is also explained in When Classic Automation Beats AI. All tables, per-article scores, runtime settings and one article compared across models are in the test report.
Frequently asked questions
Which local AI model writes the best German product descriptions?
In my test, Google's Gemma 4 31B in the 4-bit QAT build. Gemma 4 26B came close and is about four times faster. That applies to German workwear descriptions with a demanding brand prompt.
Do I need a graphics card for this?
Not a discrete one, but a lot of fast memory. The ASUS GX10 has a chip where CPU and GPU share 128 GB of memory. Gemma 4 31B in the QAT build needs about 22 GB of it.
How long do 10,000 product descriptions take locally?
With Gemma 4 31B QAT on vLLM and 8 parallel requests, about 1.6 days, including rule checks, fact checking and SEO fields.
How do I stop the AI from inventing properties?
With three layers: clean the data, use fixed rules for typical claims and run a narrowly scoped AI fact check that only corrects the affected sentence. A human spot check is still needed.