Choose a local model for document summarization

Define what a useful summary must retain

Write down the reader, the output length and the facts that matter. A three-sentence internal update and a summary of a long research paper need different checks. Start with a short source you can inspect fully.

For each test document, identify the important names and roles, quantities, dates, decisions and qualifications. Mark the difference between a forecast and a measured result. Include corrections and missing information in your test set, rather than testing only clean announcements.

Summarize the source in no more than three sentences.
Write in the same language as the source.
Preserve important numbers, roles and uncertainty.
Use only facts stated in the source. Do not invent quotations.
Return only the summary.

SOURCE:
[paste the text here]

This reusable prompt is a starting template. The recorded Swedish examples below use the original Swedish instruction, which is available with the material. Changing an instruction is a new experiment; it does not retroactively change the recorded answer.

Compare two candidates before choosing

Begin with one model that runs comfortably on your computer and a second candidate you can realistically use. Use the same source, instruction, length limit and fresh conversation for both. Record the model tag and version identifier so an updated download does not silently become a different candidate.

A parameter count or an advertised language list can help build a shortlist, but cannot establish whether a particular summary keeps the right facts. Use the language and document types of your real work.

Repeat difficult examples. Log each failure with the source passage it contradicts or omits. A single good run is useful evidence, but it is not an accuracy estimate. Keep critical factual failures separate from a preference for one writing style.

If one candidate is more concise but loses a decision or changes a number, decide whether that loss is acceptable for the intended reader. For consequential documents, review the summary before anyone acts on it.

Handle longer documents deliberately

First check whether the source fits the active context, leaving space for instructions and the answer. A model's advertised maximum context is not necessarily the context used by your local runtime. PDF extraction also needs checking: broken tables or missing pages cannot be recovered by a better summary prompt.

If you split a source, use meaningful boundaries such as sections. Keep a section identifier beside each intermediate summary, then build the final summary from those notes. Return to the original passage to verify important claims: an error in an intermediate note can survive every later step.

Test information near the beginning, middle and end. Include a correction that appears later than the claim it replaces. The short examples on this page do not establish how any model performs across a long document.

Inspect a recorded Swedish example

The source and response below are in Swedish; the interface and explanation are in English. All organisations, people and events are fictional. Choose a text, model and run to inspect the unchanged output.

Nordhamn Support testar automatisk sortering av supportärenden under sex veckor. Projektledaren Karin Ek ansvarar för försöket. I den första mätningen kategoriserades 18 av 60 ärenden fel. Försöket har inte någon bred utrullning.

Nordhamn Support testar automatisk sortering av supportärenden under sex veckor. Projektledaren Karin Ek ansvarar för försöket. I den första mätningen kategoriserades 18 av 60 ärenden fel. Försöket har inte någon bred utrullning.

Okay, let's tackle this query. The user wants a Swedish summary of the given source text in up to three sentences. They specified to keep important numbers, roles, and uncertainty. Also, no adding quotes or background.

First, I'll read the source text carefully. The key points are:

- Nordhamn Support is testing automatic sorting of support cases for six weeks.
- Project leader Karin Ek is in charge.
- In the first measurement, 18 out of 60 cases were categorized incorrectly.
- No decision on broader rollout yet.
- Written by Leo Berg, freelance writer.
- No direct quotes from Karin Ek.

Important numbers: 18/60, six weeks. Roles: Karin Ek (project leader), Leo Berg (freelance writer). Uncertainty: no decision on broader rollout.

Need to make sure the summary is in Swedish, three sentences max. Avoid adding info not in the text.

First sentence: Nordhamn Support testar automatisk sortering under sex veckor, projektledaren Karin Ek ansvarar för försöket. Then mention the first measurement result.

Second sentence: 18 av 60 ärenden kategoriserades fel i den första mätningen.

Third sentence: Inga beslut om bredare utrullning, texten skriven av Leo Berg.

Wait, the original says "Någon bred utrullning är inte beslutad." So "Ingen bred utrullning har besluts" or "Ingen bred utr." Hmm, the user wants uncertainty. The original says "Någon bred utrullning är inte beslutad." So "Ingen bred utrullning har ännu besluts" or "Ingen bred utrullning är besluten." Wait, the

Okay, let's tackle this query. The user wants a Swedish summary of the given source text in up to three sentences. They specified to keep important numbers, roles, and uncertainty. Also, no adding quotes or background, just use the info from the text.

First, I'll read the source text carefully. The key points are:

- Nordhamn Support is testing automatic sorting of support cases for six weeks.
- Project leader Karin Ek is in charge of the test.
- In the first measurement, 18 out of 60 cases were categorized incorrectly.
- No decision on broader rollout yet.
- Written by Leo Berg, freelance writer.
- No direct quotes from Karin Ek.

So, important numbers: 18/60, six weeks. Roles: Karin Ek (project leader), Leo Berg (freelance writer). Uncertainty: no decision on broader rollout.

Need to make sure the summary is in Swedish, max three sentences. Let me check the original text again to avoid missing anything.

First sentence: Nordhamas Support testar automatisk sortering under sex veckor, projektledaren Karin Ek ansvarar för försöket. Then the first measurement result: 18 av 60 ärenden kategoriserades fel. Then uncertainty: Ingen bred utrullning är besluten. Also, the author is Leo Berg, but maybe that's less critical? Wait, the user says to keep important numbers, roles, and uncertainty. The author's role might be part of the info, but the user says "bevara viktiga siffror, personers roller och osäkerhet". So roles here are Karin Ek as project leader, and Leo Berg as the writer. But the user says "endast information i källtexten". The text says "Artikeln

Nordhamn Support testar automatisk sortering av supportärenden under sex veckor, med projektledaren Karin Ek som ansvarig. I den första mätningen kategoriserades 18 av 60 ärenden fel. En bred utrullning har inte beslutats.

Nordhamn Support testar automatisk sortering av supportärenden under sex veckor, med projektledaren Karin Ek som ansvarig. I den första mätningen kategoriserades 18 av 60 ärenden fel. En bred utrullning har inte beslutats.

Bolaget Tallvik Teknik har byggt tre prototyper för att sammanfatta interna dokument. De hoppas att handläggningstiden kan minska med 20 procent, men har inte genomfört någon jämförande mätning än. En internt test med tolv dokument startar den 12 november 2026. Ingen produkt säljs till kunder, och något pris har inte offentliggjorts.

Bolaget Tallvik Teknik har byggt tre prototyper för att sammanfatta interna dokument. De hoppas att handläggningstiden kan minska med 20 procent, men har inte genomfört någon jämförande mätning än. En internt test med tolv dokument startar den 12 november 2026. Ingen produkt säljs till kunder, och något pris har inte offentliggjorts.

Okay, let's tackle this query. The user wants a Swedish summary of the given source text in up to three sentences. They specified to keep important numbers, roles, and uncertainty. Also, no adding quotes or background.

First, I need to read the source text carefully. The company Tallvik Teknik has built three prototypes to summarize internal documents. They hope to reduce handling time by 20%, but haven't done any comparative measurements yet. An internal test with twelve documents starts November 12, 2026. No product is sold to customers, and the price hasn't been publicized.

Important numbers: 3 prototyper, 20 procent, 12 dokument, 12 november 2026. Roles: Tallvik Teknik (bolaget), they are the ones building prototypes. Uncertainty: hasn't done comparative measurements, no product sold, price not publicized.

Need to make sure the summary is in Swedish, three sentences max. Let me check the original text again.

Original says: "Bolaget Tallvik Teknik har byggt tre prototyper för att sammanfatta interna dokument. Bolaget hoppas minska handläggningstiden med 20 procent, men har ännu inte genomfört någon jämförande mätning. Ett internt test med tolv dokument startar den 12 november 2026. Ingen produkt säljs till kunder och något pris har inte offentliggjorts."

So the key points:

- 3 prototyper (three prototypes)
- Minska handläggningstiden med 20% (reduce handling time by 20%)
- Ingenta jämförande mätning (no comparative measurement done yet)
- Test med

Okay, let's tackle this query. The user wants a Swedish summary of the given source text in up to three sentences. They specified to keep important numbers, roles, and uncertainty. Also, no adding quotes or background.

First, I'll read the source text carefully. The company Tallvik Teknik has built three prototypes to summarize internal documents. They hope to reduce handling time by 20%, but haven't done any comparative measurements yet. An internal test with twelve documents starts November 12, 2026. No product is sold to customers, and the price hasn't been publicized.

So, important numbers: 3 prototyper, 20 procent, 12 dokument, 12 november 2026. Roles: Bolaget (Tallvik Teknik) as the company. Uncertainty: ingen jämförande mätning genomfört, ingen produkt säljs, pris inte offentliggjorts.

Need to make sure the summary is in Swedish, three sentences max. Let me check the original text again.

Original says: "Bolaget Tallvik Teknik har byggt tre prototyper för att sammanfatta interna dokument. Bolaget hoppas minska handläggningstiden med 20 procent, men har ännu inte genomfört någon jämförande mätning. Ett internt test med tolv dokument startar den 12 november 2026. Ingen produkt säljs till kunder och något pris har inte offentliggjorts."

So the key points are:

- Tallvik Teknik har byggt 3 prototyper för att sammanfatta interna dokument.
- Hoppas minska handläggningstiden med 20 %, men ing

Tallvik Teknik har byggt tre prototyper enligt interna dokument och hoppas minska handläggningstiden med 20 procent. Ett internt test med tolv dokument startar 12 november 2026, men ingen produkt säljs ännu och inget pris har offentliggjorts.

Tallvik Teknik har byggt tre prototyper enligt interna dokument och hoppas minska handläggningstiden med 20 procent. Ett internt test med tolv dokument startar 12 november 2026, men ingen produkt säljs ännu och inget pris har offentliggjorts.

Lindåra kommun avbröjde sin pilotförsöks med en automatisk chatt den 3 oktober 2026, eftersom svar på öppettider var felaktiga. Försöket kostade 350 000 kronor och omfattade två bibliotek, inte alla verksamheter. Budgetuppgiften för försöket var felaktigt angavs som 3 500 000 kronor, och den har nu rättats. En beslut om ett nytt försök saknas.

Lindåra kommun avbröjde sin pilotförsöks med en automatisk chatt den 3 oktober 2026, eftersom svar var felaktiga. Försöket kostade 350 000 kronor och omfattade två bibliotek, inte alla verksamheter. Budgetuppgiften för försöket var felaktigt angavs som 3 500 000 kronor, och den har nu rättats. En beslut om ett nytt försök saknas.

Okay, let's tackle this query. The user wants a Swedish summary of the given source text in up to three sentences. They specified to keep important numbers, roles, and uncertainty. Also, no adding quotes or background.

First, I need to read the source text carefully. The key points are:

- Lindåra kommun stopped their pilot with automatic chat on October 3, 2026, due to incorrect opening hours responses.
- Budget was 350,000 SEK (previously 3,500,000 SEK, now corrected).
- The pilot included two libraries, not all services.
- No decision on a new trial yet.

I need to make sure the summary has the correct numbers: 350,000 kronor (not 3.5 million), two libraries, and the uncertainty about a new trial. Also, the date is October 3, 2026.

Check if the roles are clear: Lindåra kommun is the municipality, they were the ones doing the pilot. The error was in the opening hours responses.

Avoid adding any info not in the text. The user said to not invent quotes or background.

Let me draft the summary in Swedish. Three sentences max.

First sentence: Lindåra kommun avbröt pilotförsöket den 3 oktober 2026 efter felaktiga svar om öppettider. (They stopped the pilot after incorrect responses about opening hours.)

Second sentence: Budgetet var 350 000 kronor (rättat från 3 500 000 kronor). (Budget was 350k, corrected from 3.5m.)

Third sentence: Försöket omfattade två bibliotek, inte alla kommunens

Okay, let's tackle this query. The user wants a Swedish summary of the given source text in up to three sentences. They specified to keep important numbers, roles, and uncertainty. Also, no adding quotes or background.

First, I need to read the source text carefully. The key points are:

- Lindåra kommun stopped their pilot with automatic chat on October 3, 2026, due to incorrect answers about opening hours.
- Budget was 350,000 SEK (previously 3,500,000 SEK, now corrected).
- The pilot included two libraries, not all services.
- No decision on a new trial yet.

Important numbers: 350 000 kronor (corrected), 3 500 000 kronor (previous incorrect), two libraries.

Roles: Lindåra kommun (the municipality), they stopped the pilot.

Uncertainty: No decision on a new trial.

I need to make sure to mention the correction of the budget from 3.5 million to 350k. Also, the pilot was for two libraries, not all services. The reason for stopping is incorrect answers about opening hours.

Let me structure the summary in three sentences max.

First sentence: Lindåra kommun avbröt piloten den 3 oktober 2026 efter felaktiga svar om öppettider. (They stopped the pilot on Oct 3, 2026, due to incorrect answers about opening hours.)

Second sentence: Budgetet var 350 000 kronor (previously angivit som 3 500 000 kronor, nu rättad). (Budget was 350k, corrected from 3.5m.)

Third sentence: Fö

Lindåra kommun avbröt sitt pilotförsök med automatisk chatt den 3 oktober 2026 på grund av felaktiga svar om öppettider. Budgeten för försöket var 350 000 kronor, vilket tidigare felaktigt angivits som 3 500 000 kronor. Försöket gällde två bibliotek, inte hela kommunens verksamhet.

Lindåra kommun avbröt sitt pilotförsök med automatisk chatt den 3 oktober 2026 på grund av felaktiga svar om öppettider. Budgeten för försöket var 350 000 kronor, en tidigare uppgift om 3 500 000 kronor har rättats. Försöket gällde två bibliotek, inte hela kommunens verksamhet.

In the roles example, a writer should not become the project's spokesperson. In the uncertainty example, the hoped-for improvement should not become a measured gain. In the correction example, the revised budget should replace the incorrect amount.

The small Qwen model's first roles summary uses four sentences despite the three-sentence limit. That is a concrete instruction-following issue visible in the original response. It does not, by itself, establish a ranking for summary quality.

Method and reproduction

The examples were recorded locally using Ollama 0.18.2 on 2026-10-05. Each model received the same instruction and source in a fresh conversation. There were three texts for each task and two runs per text. The three models belong to one family.

Requests used temperature 0.2, context 4096 tokens, an output limit of 384 tokens and seeds 42 and 43. We requested think: false. No system prompt or schema-enforced output was used.

Summaries have not received an editorial quality score. The checklist helps a reader inspect the original response; it is not an automatic accuracy measurement.

The local qwen3:4b configuration reached the output limit in every recorded request and did not produce a complete final answer. The evidence includes those failures. Different model templates, versions or output limits need a separate test; these examples do not establish that the model itself is generally unsuitable.

ollama pull qwen3:1.7b
ollama pull qwen3:4b
ollama pull qwen3:8b
node .\run-swedish-tests.mjs

Documentation and further checks

• Ollama context length — distinguish an active window from a model's advertised capacity.
• Ollama chat API — request and response fields.
• EuroEval — a wider reference for language-model evaluation; separate from our examples.

Updated 5 October 2026. No editorial quality score or overall model winner has been assigned to these examples.