Generating AI product descriptions on your own GPU server lets you refresh the whole catalogue and translate it into many languages without a bill that grows with every token. The model works overnight on data from your PIM, and in the morning your team reviews a sample and publishes. Here is how to set it up, which models to consider and how to avoid thin, duplicated content.
This article is for online shop owners, e-commerce managers and whoever owns the product catalogue. We assume your product data lives in a PIM or the shop back office and that you sell, or want to sell, in several markets.
The problem: thousands of products, many languages, constant change
A catalogue of a few thousand products in 20 languages means tens of thousands of texts. Manufacturer descriptions are identical across hundreds of shops, so they do little for search visibility. A translation agency is accurate but expensive and slow every time the range changes. Public LLM APIs solve the speed problem, but cost grows linearly: every product, every language and every refresh means more tokens. We explain how tokens are counted in tokens in large language models.
The solution: AI product descriptions on a self-hosted model
The model receives structured data from your PIM (category, attributes, dimensions, materials, use cases) plus your brand guidelines and the keywords for each category. It returns a full set: title, short description, long description, feature bullets and meta description. The key rule in the prompt: write only from the data provided, and if an attribute is missing, leave it out rather than guess.
SEO consistency comes from per-category templates: where the keyword goes, length limits, which attributes must always appear. The result is a consistent structure across the catalogue instead of a random mix of styles.
Translation with multilingual open models
For translation and multilingual generation, test these Apache 2.0 open models (as of November 2026; the market changes monthly):
- Qwen3.8-27B: multilingual and multimodal, so it can also describe product photos;
- Gemma 4 31B and 12B: the 12B variant suits simpler, high-volume work;
- Mistral Small 4 (119B parameters): a larger model for when quality in a given language matters more than speed.
Quality varies by language. Models usually do well in the major European languages; less widely spoken ones need review by a native speaker. Prepare a glossary: brand and product-line names that stay untranslated, industry terms, units and sizes. Always validate your model choice on your own products, not on leaderboards.
Brand voice: prompting or fine-tuning
Start with the prompt: style guidelines plus a few approved descriptions as examples. That is enough for most shops. Light fine-tuning (for example LoRA) makes sense when you have hundreds or thousands of approved descriptions and prompting still misses your tone.
Overnight batch processing on a fixed-cost GPU
Description generation does not need to run in real time. A job queue starts in the evening, the inference server (for example vLLM with continuous batching) processes many products in parallel, and results wait for approval in the morning. A GPU server costs the same each month however many tokens it generates, so refreshing the whole catalogue or adding a language does not change the bill. During the day the same server can handle other work.
Whether it pays off depends on catalogue size, number of languages and how often things change. We describe how to compare GPU cost with API fees and staff time in how to calculate the ROI of an AI deployment, and the hardware access options in GPU as a Service.
Hardware you need
Memory estimate: parameters × bytes per parameter (BF16 = 2 B, 4-bit ≈ 0.5 B), plus context memory and headroom. In batch work, extra memory means more jobs in parallel.
| Scenario | Example model | Hardware |
|---|---|---|
| Smaller catalogue, a few languages | Gemma 4 12B (approx. 24 GB in BF16) or Qwen3.8-27B at 4-bit (approx. 14 GB) | GPU VPS with A10 24 GB or A40 48 GB |
| Large catalogue, many languages | Qwen3.8-27B in BF16 (approx. 56 GB) or Gemma 4 31B in BF16 (approx. 62 GB) | Server with RTX PRO 6000 96 GB or 2× A40 96 GB |
| Best quality in difficult languages | Mistral Small 4 at 4-bit (approx. 60 GB) | RTX PRO 6000 96 GB; two such cards at 8-bit |
Our ready-made GPU packages with A10 and A40 cards are a good starting point and run any model. Larger configurations are on the GPU servers page.
Integration with your PIM and shop platform
Most PIM systems and shop platforms offer an API for reading and writing products. The simplest way to build the pipeline is n8n: fetch products changed since the last run, send them to the local model, validate the output, write it back to the PIM with a “pending review” status, publish after approval. We show how to connect n8n to a local model in AI agents in n8n on a local LLM, and we deploy and host n8n for clients.
Quality control: rules and sampling
Publish AI product descriptions only after two layers of checks: automated and human.
- Automated rules: length, required attributes present, numbers matching the source (dimensions, material percentages), language detection, glossary compliance, a list of banned claims (such as health promises).
- Human sampling: review a large share of descriptions per language and category at first, then reduce the sample where errors are rare.
- Native speakers: at least a sample in every language should be checked by someone who knows that market.
Common pitfalls
- Duplicate content: an overly rigid template produces near-identical descriptions for similar products. Vary texts through attributes and use cases, not just word order.
- Invented features: the model adds an attribute that is not in the data. The “facts provided only” rule and an automatic comparison with the source help.
- Overwriting hand-crafted descriptions of your best-selling products. Flag them in the PIM as excluded from automation.
- Missing hreflang on language versions, so search engines cannot tell which version serves which market.
- Shop infrastructure: more language versions mean more pages and traffic; see when e-commerce needs dedicated cloud hosting.
FAQ: frequently asked questions
Does Google penalise AI product descriptions?
According to Google’s public guidance, it assesses the quality and usefulness of content, not the tool used to write it. The problem is mass-produced, duplicated text that gives shoppers nothing.
Which open-source model is best for translating product descriptions?
There is no single answer for all languages. Test Qwen3.8-27B, Gemma 4 and Mistral Small 4 on a few dozen products in each target language.
How much does it cost to generate product descriptions on your own GPU?
A fixed monthly fee for the server, regardless of how many texts you generate, plus your team’s time for quality control. Our ready-made GPU packages are a good reference point.
Do I need to fine-tune a model for my brand voice?
Usually not. Good guidelines and examples in the prompt are enough; consider fine-tuning once you have a large set of approved descriptions.
Can this connect to my online shop?
Yes, if your platform or PIM has an API. We usually build the integration in n8n.
We can run a model for product descriptions and translations on a GPU server in a Tier III, ISO 27001-certified data center in Poland, with 24/7 administration. See our GPU packages or contact us and we will help you work out whether it pays off for your catalogue.