Torad Labs

Torad Labs is a research lab in Austin, Texas. We train small models for specific tasks and build the tools to quantize them and run them on your own hardware.

What we work on

• Our RL post-training engine scores each span of a response separately, trains in phases that get harder over time, and uses a Fisher-SVD optimizer in place of AdamW.
• Our inference engine is written in Rust and CUDA. It runs the whole model as one kernel, compiles it through its own IR, and uses no NVIDIA libraries.
• We design low-bit weight formats that fit large models on consumer and edge GPUs and keep quality close to bf16. Our llama.cpp fork has CUDA kernels for these formats on RTX 50-series cards.
• Our research covers continued pretraining, reinforcement learning, world models, Gaussian splats, and structural connections across domains.

torad.ai · GitHub