Goodbye paid translation: building localisation on open source and an LLM (Прощай, платный перевод: собираем локализацию на open source и LLM)
B · Useful, limited YADRO technical-documentation engineer Ruslan Tashmetov documents an on-prem localisation stack with tool choices, prompts and a per-word cost table; a useful build log, but no translation-quality measurement and a cost estimate that leaves out most operating costs.
Summary based on the full text·Original: about 10 minutes to read
In short
A YADRO company blog post on Habr (2 November 2026) by Ruslan Tashmetov. The technical-documentation team dropped Smartcat, a cloud CAT tool that bills per uploaded word whether new or already translated, for an on-prem stack nicknamed "Tolmach": Serge pulls strings from Git into PO files, Weblate is the translator interface, and Dify sends suggestion requests to a local Qwen3.6-35B-A3B model. The author estimates the in-house cost at no more than 0.40 RUB per word against about 2.5 RUB for Smartcat.
Key numbers
- about 0.40 RUB vs about 2.5 RUBcost per word on the in-house cluster in the worst case (one-word segment) vs current Smartcat rates
- about six timeshow much cheaper the in-house stack is than Smartcat, by the author's worst-case comparison
- 0.0048 RUBestimated cost per word for a 250-word segment
- 5 million RUB, 5 years, 30-40%GPU price, service life and load assumed in the cost estimate
- about 0.63 secondstime to return one machine-translation suggestion
As reported by the source.
Russian·Web page·opens habr.com
What it reports
- Requirements: all text stays on YADRO servers, the LLM and its parameters can be changed, and the tool has a web or client-server interface.
- Zing, the Evernote translation server built to pair with Serge, was rejected because Evernote stopped maintaining it and was looking for a new maintainer. Weblate was chosen because it is actively developed and has paid support.
- All formats (DITA, config files, PDF, source code) are converted to PO files, so one pipeline covers them.
- Formatting rules in the prompt hurt translation quality. The model kept writing the letter "ё" despite instructions, so a code step now replaces it after generation. Structured JSON output with a single field stopped the model adding chatty preambles.
- Known issues: the model can change meaning, ignore terminology and break formatting, so a technical writer still proofreads every document.
- Cost per word falls with segment length: 0.40 RUB for 1 word, 0.0112 for 50, 0.0064 for 125, 0.0048 for 250.
- Upload to Weblate was scripted against its REST API because the standard push and pull commands did not work.
How they got it
A first-person build account with screenshots, prompts and configuration. The cost table assumes a 5 million RUB GPU, a five-year life and steady 30-40% load.
Limits
- No measurement of translation quality, edit distance or reviewer time before and after.
- The cost estimate excludes the server, electricity, engineers' salaries and other operating costs; the author says real cost will be higher.
- The "six times cheaper" figure compares the in-house worst case with Smartcat's per-word list rate, not total cost of ownership. Part one of two; the deployment guide is promised later.
Why it matters for companies
Per-word cloud pricing that charges for repeated text is a real cost driver for documentation-heavy firms, and open-source CAT tools plus a local model are now a workable replacement where data must stay in-house. Two lessons transfer to other LLM pipelines: move deterministic formatting rules out of the prompt and into code, and use a structured output schema to keep responses clean. Budget for the human proofreading step; the post keeps it.
Key passages in translation
Russian original
Тариф считается по сетке за каждое загруженное в него слово (неважно, новое оно или уже переведенное), а проектов у нас сотни.
The tariff is charged for every word uploaded (whether new or already translated), and we have hundreds of projects.
Russian original
Поэтому все, что можно сделать кодом, нужно и правильно делать кодом.
So everything that can be done in code should be done in code.
Russian original
В него не входят затраты на сервер, электроэнергию, зарплаты обслуживающих инженеров и другие операционные расходы, поэтому реальная себестоимость будет выше.
It does not include the server, electricity, the salaries of support engineers or other operating costs, so the real cost will be higher.
Summary based on the full text. Read the original for the full argument and data.Read the original on habr.com
- Tags
- AI enablement