GPT-SW3
No estimate
No hardware requirements for this model
The weights for this model have not been published, so it cannot be downloaded or run on your own hardware at any size. It is reachable only through its provider, and no graphics card changes that.
On record
Full specification
Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.
Origin
Who built this model, where, and when it was published.
- Organisation
- AI Sweden,RISE
- Organisation type
- Academia
- Country
- Sweden
- Published
- 25 June 2022
- Authors
- Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Ohman, Fredrik Carlsson, Magnus Sahlgren
What it does
The problem areas the model was built for. A model can carry several of each.
- Domain
- Language
- Task
- Language modeling
- Approach
- Self-supervised learning
Size
How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.
- Parameters
- 3.5B
- Training data
- 101,711,872,000 tokens
- Epochs
- 5
100GB Swedish corpus, assume Swedish has similar 167M words per GB as German. 100*167e6 = 1.67e10 "The model is trained for 97000 steps with a batch size of 512 using the autoregressive next-step prediction objective" "Sequence length 2,048" 97000*512*2048=101711872000 tokens = 102B tokens 100*167000000/0,75=22266666666 tokens = 22B tokens (then I assume multiple epochs)
Training compute
The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.
- Training compute
- 1.7 × 10²¹ FLOP
- How it was established
- Hardware,Operation counting
From section 4: "Training was performed on GPU resources from the Berzelius Superpod, which is currently the fastest super computer in Sweden, equipped with 60 Nvidia DGX A100 servers, each of which consists of 8 Nvidia A100 GPUs with 320 GB Total GPU memory. Our training process took 2.5 days utilizing 16 of the DGX A100 servers (in total 128 GPUs)." 2.5*24*60**2 * 128 * 1.56E+14 * 0.3 = 1.3e21 6 FLOP / parameter / token * 3.5 * 10^9 parameters * 97000 steps * 512 samples per batch * 2048 tok…
The training run
What it physically took to train: which chips, how many, for how long, and what that drew from the wall.
- Training hardware
- NVIDIA A100
- Wall-clock time
- 60 hours
"Our training process took 2.5 days utilizing 16 of the DGX A100 servers (in total 128 GPUs)."
Availability
Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.
- Weights
- Closed — provider access only
- Model access
- Unreleased
- Training code
- Unreleased
they released many other models from the series but not 3.5B one https://huggingface.co/AI-Sweden-Models/gpt-sw3-40b/blob/main/LICENSE
How it is classified
Labels the source dataset applies when tracking notable models, and how confident it is in the entry.
- Record confidence
- Confident
- Citations
- 20
Sources
Where this record came from and when it was last checked.
- Reference
- Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish
- Last updated
- 28 November 2025
What the numbers mean
What this model is
GPT-SW3 was published by AI Sweden,RISE, in the country recorded as Sweden, during June 2022. The category the publisher falls under is academia.
It works in the domain of Language, and is recorded as performing the task of language modeling.
Because the weights are not available, none of the hardware figures elsewhere on this site apply to it.
Training and provenance
Training it took a computation budget of roughly 1.7 × 10²¹ FLOP, on hardware recorded as NVIDIA A100. That figure measures what producing the model cost, and has no bearing on how fast it answers.
It was trained on a corpus of about 101,711,872,000 tokens of text.
Answers
GPT-SW3 — common questions
GPT-SW3— when was it released?
It was published in June 2022. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.
GPT-SW3— what is it used for?
It works in the domain of Language, and is recorded as handling the task of language modeling. A model can carry several of each, so these are the areas it was built for rather than a limit on what it will attempt.
GPT-SW3— how much compute was used to train it?
Training consumed around 1.7 × 10²¹ FLOP, on hardware recorded as NVIDIA A100. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.
GPT-SW3— what GPU do I need to run it?
None. This is a closed model — its weights were never published, so it cannot be downloaded or run on your own hardware at any price. It is reachable only through its provider.
GPT-SW3— is it open source?
No. Its weights have not been published, so it exists only as a service controlled by its owner.
GPT-SW3— how many parameters does it have?
It has a parameter count of 3.5B. That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.
GPT-SW3— who created it?
It was published by AI Sweden,RISE, based in Sweden, an organisation categorised as academia.
The other direction
Looking at it from the other side?
This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.