Completed research milestone · v1
ATG Base 64M
Built from scratch in Sri Lanka.
64.3M parameters · 600M training tokens
English · Sinhala · Tamil
ATG Base 64M v1
Building the stack,
not just an application.
An independently trained 64.3-million-parameter trilingual ATG base model, built from random initialization using ATG's own decoder architecture and trilingual tokenizer. Base 64M completed 600 million tokens of English, Sinhala and Tamil pretraining and represents the next major milestone in OpenATG's compact-model research programme.
Model and training
- Parameters
- 64,307,712
- Architecture
- ATG Decoder
- Tokenizer
- ATG Trilingual 16K v2
- Initialization
- Trained from random initialization
- Training content tokens
- 600,000,000
- Optimizer updates
- 14,717
- English
- 240,000,000 tokens
- Sinhala
- 180,000,000 tokens
- Tamil
- 180,000,000 tokens
- Training completed
- September 2026
- Final recorded training loss
- 3.393438551641386
- Model type
- Base model · Not instruction-tuned
Post-training validation
Post-training validation has been completed internally. Public benchmark reporting is undergoing provenance and distribution review.
Training loss alone does not establish factual accuracy, reasoning capability or generation quality.
Technical Notes
Why Base 64M matters
OpenATG is building the model stack, not simply fine-tuning an existing international foundation model.
ATG Base 64M began from random initialization and uses an ATG-authored decoder architecture together with an independently trained English-Sinhala-Tamil tokenizer.
The milestone demonstrates the complete path from corpus preparation and tokenization through model initialization, GPU training, checkpointing and post-training evaluation.
Base 64M is a research base model rather than a finished assistant. Generation quality, factual reliability and repetition remain active research areas.
Building ATG in public
A model history.
An ongoing research programme.
OpenATG preserves major model generations so collaborators, researchers and investors can follow the evolution of the architecture, training pipeline, data strategy and evaluation methodology over time.
01 / Earlier generation
ATG Base 32M
Early scaling milestone
View the existing release →02 / Completed milestone
ATG Base 64M v1
64.3M parameters
600M-token trilingual pretraining completed
September 202603 / Looking ahead
Continuing research
Architecture, data and evaluation.