title: KeelCore
emoji: ⚓
colorFrom: blue
colorTo: yellow
sdk: static
pinned: false
KeelCore is one compact base model plus many small LoRA adapters. We pretrain the base once. A new capability then arrives as a small adapter, not a new model, so you download the base once and add only the adapters you need.
KeelCore-750M · Technical report · GitHub · Contact
Models
| Model | Status | Details |
|---|---|---|
| KeelCore‑750M | Base model | 752M parameters, encoder–decoder, 2,048-token context. English, code and command line. |
Benchmarks, including where we trail models trained on larger budgets, are on the model card.
Why this approach
- One base. A single pretrained encoder–decoder with a tokenizer built for English prose, source code and command-line text.
- Small adapters. A rank-16 LoRA adapter trains about 18.3M parameters, 2.4% of the base, while the base weights stay frozen.
- Runs on your device. A weight-only int8 ONNX export (0.76 GB) runs in a web browser through ONNX Runtime Web. Because that runtime could not load LoRA adapters, we contributed the support ourselves: a
LoraAdapterAPI and anactiveLoraAdaptersrun option, merged in October 2026 (microsoft/onnxruntime#32886).
License
Models are released under the KeelCore License, Version 1.0: personal use through the official EchoMeBetter extension and noncommercial research, which requires citing the project. Commercial use, use by an organization and any use outside the extension need a separate written agreement. Adapters follow the same terms unless their repository says otherwise.
Contact
Saravanan A R · LinkedIn · GitHub
Commercial licensing and partnerships: sararavi.ar@gmail.com