Independent software studio
Useful software, and a small AI of our own.
Avenpa makes practical tools for the small jobs that deserve a clear answer, and is training its own small language model, Wren v1, alongside them.
- 01Trained from scratch
- 02Runs on CPU, no GPU
- 03Korean and English
Products
Software from the studio. One model in training, one tool site live.
Wren v1
Inside Wren v1
A small model, built in the open.
Ask about the model
What Wren v1 is, what it is trained on and where it stands today.
ExploreTokens
Korean and English text broken into the pieces the model reads.
ExploreTraining
A loss curve that falls as the model sees more tokens. Illustrative.
ExploreScale
900M training tokens, around 340 tokens a second, four CPU threads.
ExploreTrain
A training run on four CPU cores, resumable from its last checkpoint.
ExploreInside Avenpa Tool
Four kinds of small job. Each one finished in a tab.
Money and payroll
10 tools. Net salary, severance, overtime and year-end tax, worked out under Korean labour rules that are dated by year and covered by tests.
ExploreDates and time
8 tools. Day counts, D-days, ages and instalments, counted against the real calendar instead of an average month.
ExploreImages, on your device
7 tools. Resize, merge, censor and compress photos without uploading a single pixel. The browser does the work.
ExploreNumbers and text
8 tools. Percentages, unit ratios, character counts and QR codes. The arithmetic you would otherwise do twice.
Explore1.24B
Tokens in the Wren v1 corpus
900M
Tokens in the training run
0
GPUs used
33
Tools live on tool.avenpa.com
Company
A small studio, on purpose. The person who trains a model stays to explain it, test it honestly and make it better.
About AvenpaResearch
Notes from the training room
Wren v1: a small language model on four CPU cores
A dense language model for Korean and English, pretrained from scratch on a CPU-only server. It is still training.
Rules that change every January
Rates that move on dates we do not control, kept in one dated module so the calculators cannot disagree with each other.
A calculator that is quietly wrong is worse than none
Money pages assert the thresholds and the rounding, and read the rates out of the source instead of carrying a second copy of them.
How we build
Small tools, held to a high bar.
Build
Explore01
Fast by default
Every page is prerendered and served from the edge. There is no spinner between you and the answer.
02
Nothing to sign up for
Tools open for anyone with the link. An account exists only if you want to save something you made.
03
Numbers carry tests
Rates and thresholds are asserted in the test run, and each page says which year they apply to.
Have a use for a small model?
Tell us what you would want it to do. A person at the studio will get back to you.