Release · 11 November 2026
Introducing A·we 1 (beta)
A·we 1 is JAKASII's language model for Garo (A·chikku), trained on Garo data we built and reviewed with fluent speakers.
- Trained specifically for Garo, a language almost absent from today's AI training data
- Built on a JAKASII Garo corpus of 19M+ tokens and thousands of speaker-reviewed examples
- Perplexity on unseen Garo text cut from 95 to about 6
- API and public benchmark on 18 November
Why we built it
Garo is spoken by communities across Meghalaya, Assam and Bangladesh, yet it remains largely unsupported by today's AI systems. Existing models frequently misidentify the language and produce inaccurate vocabulary. JAKASII believes Garo speakers deserve AI that works in their own language. A·we 1 is the first step in that commitment.
How it was built
Data. JAKASII has built a proprietary Garo corpus of nearly 31,000 documents, alongside thousands of instruction examples spanning question answering, translation and dialogue, each reviewed by fluent speakers. A held-out evaluation set is maintained separately and is never used in training.
Training. A·we 1 is trained in two stages: large-scale language modelling on Garo text, followed by instruction tuning on reviewed data.
Evaluation
| Measure | Before | A·we 1 |
|---|---|---|
| Perplexity on unseen Garo text (lower is better) | 95.1 | 6.3 |
| Translation quality (chrF) | 18.8 | 27.3 |
| Overall answer quality (chrF), vs our earlier checkpoint | 18.2 | 24.6 |
The full method and per-task scores will be published with the benchmark on 18 November.
Availability
A·we 1 is in beta. The A·we API opens on 18 November. In JAKASII Chat, A·we 1 works alongside larger hosted models. Try JAKASII Chat
Current limitations
As a beta release, A·we 1 is still maturing in open-ended conversation and factual accuracy. We are actively broadening its coverage of conversational Garo, and these areas will improve with each release.
What's next
JAKASII is expanding its reviewed conversational dataset in partnership with fluent Garo speakers. Progress will be reported transparently through our benchmark, beginning 18 November.