Foundation models

Diffusion language models that write a whole answer at once, sold by API.

A different inference architecture — diffusion rather than autoregression — which changes the latency and cost-per-task arithmetic for any product that calls a model many times inside a single user interaction.

A Tokyo lab selling routed multi-model access and Japanese-language models.

They sell orchestration over open weights as a route to frontier scores at lower cost — the routing layer above the model, priced per token like a model, which is where the buy-versus-route decision actually gets settled.