Kedy Atlas

Research

Building efficient AI,
openly.

We publish our methods, benchmarks, and findings so the research community can build on what we learn.

01

Efficient model architectures

Exploring how 1B-parameter models can match the practical performance of much larger systems through architecture choices, distillation, and targeted training.

02

Domain-specific fine-tuning

Developing principled approaches to fine-tuning general-purpose checkpoints for scientific, coding, and conversational tasks without catastrophic forgetting.

03

Evaluation & benchmarking

Designing evaluation frameworks that reflect real-world task difficulty rather than narrow benchmark saturation — and releasing them for open use.

04

Inference optimization

Quantization, pruning, and runtime techniques that make sub-2B models viable for CPU-only deployments without meaningful quality degradation.

05

Alignment at small scale

Studying how RLHF and constitutional approaches transfer to smaller models, and where smaller-scale alignment diverges from large-model behavior.

06

Open data & reproducibility

Curating and releasing the training data mixes, evaluation sets, and pre-processing pipelines that underpin Kedy's models.

Collaborate with us.

We're interested in working with academic researchers and institutions. If you have a project that aligns with our work, reach out.