Models and datasets

Compact medical and healthcare language models from Tether Data's AI Research, built for edge deployment. Across seven medical benchmarks the 1.7B model outperforms Google's MedGemma 4B by over eleven points despite being less than half its size, and on HealthBench Hard it beats the roughly sixteen-times-larger MedGemma 27B. The 4B variant outperforms MedGemma 27B on the same seven benchmarks. Four-bit quantization holds accuracy within about a point of full precision while cutting disk footprint by roughly 69%.

Synthetic pre-training datasets for STEM and logic. Genesis I shipped 41 billion text tokens across nineteen educational fields in October 2025; Genesis II followed, bringing the combined total to 148 billion. The stated aim is to give independent groups the training data to build competitive models without Big Tech's corpus.

A fork of llama.cpp and the engine under everything else. Hardware-agnostic across NVIDIA, AMD, Intel, Mali, Adreno, and Apple silicon, switching between Vulkan, CUDA, and ROCm as appropriate. Its Dynamic Tiling Algorithm segments large matrix operations to work around mobile memory limits, and it is the first framework to enable LLM fine-tuning directly on mobile devices — combining BitNet and LoRA to fine-tune a billion-parameter model on a Galaxy S25 in about eighty minutes.