Two starting points, different units of comparison.
Viking 33B is a specific model size; GPT-SW3 is a family. Comparing their names alone does not establish which checkpoint will work better for your application. This guide covers two documented projects in our directory, rather than a complete or ranked list of Nordic AI.
| Question | Viking 33B | GPT-SW3 |
|---|---|---|
| Who developed it? | Silo AI and TurkuNLP | AI Sweden, RISE and WASP WARA Media & Language |
| Language focus | English, Finnish, Swedish, Norwegian, Danish, Icelandic and code | Nordic languages, with a primary focus on Swedish |
| What to select | The exact Viking checkpoint and version | A specific family checkpoint and version |
| Before deployment | Read the checkpoint model card and licence | Read the checkpoint model card and licence; add application-specific safeguards |
AMD’s account describes Viking training on LUMI in Finland. AI Sweden identifies the GPT-SW3 collaboration and warns that its models can produce incorrect information. These are provider accounts; Frontier AI has not reproduced their training or benchmark results.
Source: AMD / Viking 33B ↗ · Source: AI Sweden / Models & Resources ↗
Start with a task you can check.
The following method is our editorial guidance. Write down the language, intended reader and output you need. A Swedish support reply, a Finnish summary and an Icelandic search query are different tests. Include dialects, local names, spelling variations and mixed-language examples relevant to your users.
Build a small evaluation set.
- Collect 20–50 representative examples using public, synthetic or appropriately authorised material.
- Write an expected answer or a short scoring rubric before running the models.
- Include examples where the correct response is uncertainty or refusal, plus misleading instructions and irrelevant context.
- Use the same prompt and generation settings where possible. Record differences where formats or context limits prevent this.
- Ask a fluent reviewer to score accuracy, completeness and natural language quality. Keep failure examples.
Record the conditions, not just the result.
For each run, record the checkpoint identifier, revision, prompt, date, settings, runtime and hardware or hosting provider. Measure latency and resource use in your own environment. A model’s name or parameter count alone does not establish speed, cost or suitability.
A reusable review worksheet
- Task and language: What exactly must the system do?
- Evidence: Which source supports each model claim?
- Version: Which checkpoint and revision did you test?
- Terms: What licence applies to that checkpoint and your intended use?
- Results: Which examples passed, failed or required human review?
- Decision: What remains unresolved before a pilot?
Read access terms separately from capabilities.
Being able to download weights does not establish unrestricted permission for every use. Follow the licence attached to the exact checkpoint, including any additional conditions. Likewise, Nordic development or training infrastructure alone does not establish where a hosted service will process your data.
Keep the comparison useful.
Use our comparison tool to organise documented facts, then follow the model cards before testing. Missing context-window or licence details in our directory mean we have not recorded a verified value; they do not mean the model lacks that feature. This guide makes no claim that either project is the latest or best available option.
Found a missing source or changed checkpoint? Send a correction →