Arc 1.0
Our first model. A compact assistant that reads and writes Burmese, small enough to run on a laptop.
- Expected
- December 2026
- Parameters
- 4B effective
- Weights
- Open source
Building the AI Frontier of Myanmar
Arckara is an AI research company building language models for Burmese.
Open source · Expected December 2026
Introducing Arc 1.0
Our first model is a compact assistant that reads and writes Burmese. It is built on Gemma 4 and small enough to run on a laptop.
Mission
Burmese makes up a tiny share of the text that language models learn from. Arckara trains models on Burmese first. We keep them small enough to run on a laptop, and we publish how every version measures up.
Models
Explore our models
Each release comes with a model card and the evaluation results behind it.
Arc 2.0
Our next model. Details will follow once Arc 1.0 is out.
- Expected
- To be announced
- Parameters
- To be announced
- Weights
- To be announced
Arc 3.0
Further out. Details when there is something to measure.
- Expected
- To be announced
- Parameters
- To be announced
- Weights
- To be announced
Waitlist
Be the first to try Arc
Leave your email and we will write when each Arc model is released. Nothing else.
Research
How we build and measure Arc
Arc 1.0 is in training for release in December 2026. Beside each step is what the release is planned to reach. The benchmark below is measured by us.
- 01
Data
20,000Burmese instruction rows planned for Arc 1.0, about 7.5 million tokens
An open teacher model, Gemma 4 26B, writes Burmese instructions and answers. Written rules remove broken, padded and wrong rows, and a native speaker reviews samples from every dataset. No closed model writes or grades our data.
- 02
Training
100+NVIDIA T4 GPU hours planned for Arc 1.0, from writing the data to the last evaluation
Each run trains a LoRA adapter on Gemma 4 E4B on NVIDIA T4 GPUs, for two passes, with rows held back to check that it learns rather than memorises. The plan to release grows the data from 5,000 to 20,000 rows over four runs, and keeps the best one.
- 03
Evaluation
4,400+test questions for every version, plus blind native review
Every version is scored on the full test sets of Belebele for reading, FLORES-200 for translation in both directions, and GSM8K and HumanEval to catch damage to math and code. Native speaker review then compares its answers blind against the base model.
Benchmark
Choosing a base model
Reading comprehension in Burmese on Belebele, 900 questions, for three open models before any training. Translation scores on FLORES-200 were close for all three, so comprehension decided the choice.
Accuracy, percent. Measured by Arckara.
Company
Arckara comes from အက္ခရာ, akkhara, the Burmese word for alphabet. The arc is for the round letters of the script.
Arckara is an independent AI research company, founded in 2026. We would like to hear from researchers, and from Burmese speakers who want to help evaluate our models.
Contact