Qwen3.8 matches Gemini Flash on facts in a German language test

Since August, Hetzner, the German hosting provider, has been offering a second language model free of charge next to Qwen3.6, namely Qwen3.8-27B. On 21 September 2026 I tested whether it is good enough for German office work or whether Gemini Flash, at roughly half a US cent per task, is worth paying for. I sent five tasks of the kind every small business deals with to four models, five times each, 100 runs in total.

The trigger is an old finding. In my first German test in July, Qwen3.6 dropped a Chinese word into the middle of a formal German business email in four of seven runs. In the re-measurement in August, Qwen3.8 wrote the same email cleanly seven times. Seven emails are not a working week, though, so this time I ran a broader test with a paid model as the benchmark.

How I tested

The five tasks cover typical desk work and can still be checked objectively:

1. Rewrite a casual message as a formal German business email, using exactly the same prompt as in July.
2. Answer a customer enquiry to a carpentry business, with five given facts about price, on-site measurement, delivery time and travel costs.
3. Summarise meeting minutes of just under 300 words in at most 120 words without losing decisions, amounts or responsibilities.
4. Extract an invoice with three line items as JSON, a structured record that software can process directly.
5. Write an appointment confirmation for a physiotherapy practice in which date, weekday and time from the request have to be right.

Companies, names and numbers are made up. Qwen3.8-27B and Qwen3.6-35B ran on Hetzner’s inference API, Gemini 3.7 Flash and Gemini 3.8 Flash on Google’s API. I called the Qwen models without their reasoning mode, because for tasks this short it only costs time, and the Gemini models with Google’s default settings. I automatically checked whether every given detail appears in the output (215 checks per model), whether Chinese, Japanese or Korean characters show up and whether the JSON is valid. Then I read all 100 texts myself and noted content errors, language errors and breaks in style.

Results at a glance

Metric	Qwen3.8 (Hetzner)	Qwen3.6 (Hetzner)	Gemini 3.7 Flash	Gemini 3.8 Flash
Given details carried over correctly	215 of 215	210 of 215	215 of 215	215 of 215
Runs with Chinese characters	0 of 25	1 of 25	0 of 25	0 of 25
Invented or distorted details	1	5	0	1
JSON valid without clean-up	4 of 5	5 of 5	3 of 5	4 of 5
Response time per task (median)	48 s	2.8 s	4.4 s	4.5 s
Cost for all 25 tasks	free	free	0.12 US dollars	0.14 US dollars

On accuracy, Qwen3.8 is level with both Gemini models. It carried over every given detail, changed no number and did not write a single Chinese character. The difference is response time.

Business email and customer reply

The July task is the direct comparison. In its second run Qwen3.6 again wrote “bei einer预计igten Bearbeitungsdauer”, with 预计, the Chinese word for “expected”, in the same place as in July and August. In the first run an English word slipped through instead, “die corresponding Texte”. On top of that came a garbled sentence and a casual “Passt das?” (“Does that work?”) at the end of a formal email.

Qwen3.8 had none of these problems. The five emails are well built, with the correct formal salutation, every detail and a fitting closing. Once, the placeholder at the end read “[Your Name]”, English in a German email. Anyone proofreading spots that immediately, and it changes nothing about the content.

Gemini delivered equally clean emails, with one exception that has to do with the prompt. The July prompt is written without German umlauts, so “hoefliche” instead of “höfliche”. In one of five runs Gemini 3.7 Flash copied that and wrote the entire email with “ue” and “oe” in place of ü and ö, down to “Mit freundlichen Gruessen”. No other model did this.

Photo: Miguel Á. Padriñán / Pexels

The reply to the carpentry customer showed a difference in style. The Gemini models followed the instruction not to invent anything so closely that they usually also left out the thanks for the enquiry. Gemini 3.8 Flash opened four of five replies directly with the price, Gemini 3.7 Flash three of five. Both Qwen models thanked the customer in every run. Qwen3.6, however, once slipped in perspective and charged the customer for “your travel to Kassel”, even though it is the carpenter who travels.

The smaller model makes things up when summarising

The meeting minutes were where Qwen3.6 fell behind most clearly. Three of five summaries contained something that is not in the minutes. Once a named partner had supposedly approved the budget, once the meeting started “at about 9 a.m.” and once the model reversed a request. The minutes ask that the coffee machine should no longer be descaled right before the meeting, the summary said exactly that was to be done before the meeting from now on. In two further runs individual details were missing, such as the monthly costs.

Qwen3.8 stuck to the facts. All twelve checked details were in all five summaries. Two runs clearly overshot the 120-word limit, with 144 and 145 words. Once, future tasks also slipped into the past tense, as if the client mailing had already gone out. The Gemini models stayed within the word limit every time and remained factually correct.

All four get every amount on the invoice right

Extracting the invoice produced not a single content error. In all five runs, all four models correctly carried over invoice number, date, due date, the three line items, net amount, VAT, gross amount and IBAN. German notation such as “0,12 €” and “14.09.2026” was cleanly converted into numbers and the requested date format.

The difference was in the packaging. Although the prompt explicitly asked for the JSON only, Gemini 3.7 Flash wrapped the data in a Markdown code block in two runs, Gemini 3.8 Flash and Qwen3.8 in one each. Software that reads the reply directly fails on that. Qwen3.6 returned pure JSON five times. If you automate this kind of extraction, strip the wrapper in code first or use Gemini’s built-in JSON mode, which I deliberately left off here.

Photo: Kaboompics / Pexels

Gemini copies the appointment date too literally

All four models solved the appointment confirmation correctly in terms of content. Each confirmed the new appointment rather than the cancelled one and carried over weekday, time, duration and address. Qwen3.6 once mixed up the roles and wrote the equivalent of “you can expect us at our practice”.

What stood out was a grammar mistake that appeared almost only with Gemini. The patient wrote “Donnerstag, der 8. Oktober” (Thursday, 8 October), and the Gemini models dropped that phrase into their sentence unchanged. The result was “am Donnerstag, der 8. Oktober”, while according to the Duden, the standard German dictionary, the date after “am” takes the dative or accusative, so “dem” or “den”. Gemini 3.7 Flash did this in five of five runs, Gemini 3.8 Flash in four, Qwen3.8 in one and Qwen3.6 in none. Presumably Gemini took the instruction to copy the date exactly more literally than intended. A small error, but it shows that even a paid model needs someone to proofread.

The catch is waiting time

On content, then, there is little against Qwen3.8. The problem is time. Between 16:19 and 16:42 German time, the model needed a median of 48 seconds per task, the fastest was done after 7 seconds, the slowest after 2 minutes and 11 seconds. A large part of that was spent in the queue, with a median of 24 seconds until the first piece of text arrived. Qwen3.6 handled the same tasks in a median of 2.8 seconds, the Gemini models in 4.4 and 4.5 seconds, even though they reason before every answer.

Just before that it was even slower. During my speed measurement between 16:05 and 16:19, Qwen3.8 wrote only 4 to 7 tokens per second, compared with 28 in August. An explanation of 400 words took up to four and a half minutes including the wait. Load clearly fluctuates a lot. That is exactly what makes the model hard to plan with for anything where a person is waiting for the answer.

Photo: Wayne Jackson / Pexels

For background tasks it hardly matters. Nobody notices whether an invoice lands in the system after five seconds or after a minute, as long as the workflow is not waiting for it. I describe how to plug Hetzner’s models into such a workflow in n8n with Hetzner’s free AI. Keep the limit of ten requests per minute per key in mind, so calls have to be paced.

What Gemini Flash costs for this

All 25 tasks together cost 0.12 US dollars with Gemini 3.7 Flash and 0.14 US dollars with Gemini 3.8 Flash, roughly half a cent per task. Extrapolated, that is about 5 to 5.50 dollars for 1,000 tasks of this kind. More than 80 percent of it went to reasoning tokens, which Google bills at the output price even though they never appear in the result. How far a lower reasoning level brings that down is something I measured in what Gemini Flash costs per task.

Two more things belong in this calculation. Google’s prices of 0.75 dollars per million input tokens and 3.75 dollars per million output tokens only apply until 31 December 2026 and double from 1 January 2027 (as of 21 September 2026). And Qwen on Hetzner is free because it is an experiment, with no availability commitment and no data processing agreement as the GDPR requires for personal data. Customer data with personal information does not belong there, however well the model writes.

My recommendation per task

For invoices and other documents that are processed further as records, Qwen3.8 on Hetzner is entirely sufficient, provided the data contains no personal information. Even the faster Qwen3.6 worked flawlessly here. What matters is that software validates the output before it is used.

Summaries without time pressure, for example of internal minutes, can be handled by Qwen3.8 as long as someone keeps an eye on the length. I would not use Qwen3.6 for this, because three of five summaries contained something that was not in the original.

For emails to customers, the deciding question is whether someone is waiting. Qwen3.8 writes good German for drafts that a person reads before sending anyway. If the reply has to arrive within seconds, for example in a chat on your website, Gemini Flash is the safe choice at about half a cent per reply.

Appointment confirmations and similar standard texts are within reach of all four models. It is worth asking first whether a fixed text template with the date filled in would do, since that needs no language model at all and adds no grammar mistakes.

If you want to know which writing tasks in your business a free model can handle and where Gemini is worth it, get in touch and I will test it with your own tasks.