Together AI said its research team is focusing on scaling research capacity to enable abundant intelligence, shifting from Static Inference to Polymorphic Inference in response to the velocity of new model releases, diversity of model architectures, fast-changing workloads, and fast-changing hardware platforms. The company described Polymorphic Inference as inference in “many shapes” and a response to industry changes defined by converging trends: increasing velocity of model releases, with 80% of traffic on OpenRouter going to models released in the last three months; increasing diversity of model architectures, with all of the top-ten most popular OSS models in use today using different architectures; traffic shapes that continuously evolve as it serves new users, organizations, and geographies across workload shapes and SLAs, with product or algorithmic changes in harnesses and post-training of open-source models potentially changing traffic shapes; and hardware changing faster and more frequently, from B200 to GB300, with the first Vera Rubin deployments approaching, alongside CPU-level optimizations, advanced storage systems, and alternative inference chips. Together AI contrasted Polymorphic Inference with the prior era of Static Inference, defined by closed models, traditional Transformer architectures, single-user, and single-hardware-platform inference. It said it is continuously using AI agents to optimize their own inference with Continuous Compilation.