The LLM Data Company is creating specific intelligence for the jagged frontier.

Frontier models can do competition math, operate a computer and write sonnets. Invariant across all of these is a general talent: the model understands natural language and has a sense of the natural world.

Every model is a purported generalist, yet each model's vocational training comes to bear when it’s pushed. Users already reach for different models by task: one for coding, one for deep research, one for writing, and so on.

This speciation is no accident. Post-training curricula are hand-crafted and trade-offs are made at the policy level. If you want your model to follow instructions to perform on SWE-Bench, you necessarily accept some level of sycophancy. In certain domains, like medicine, this tradeoff is unacceptable.

Most models have opted for coding and tool-use on the Pareto frontier. The LLM Data Company is serving the underserved domains where models must handle ambiguity, push back, and resist sycophancy.

We are training models that push at the shortest ends of the jagged frontier. Today, we are training the Kos series of medical models, starting with Kos-1 Lite, SOTA on HealthBench Hard.