March 2025

The design made additional reasoning a choice within one model, with consequences for response time and cost.

Sharing these components exposed the engineering behind model training and inference that benchmarks alone do not describe.

The release gave developers access to the models themselves, widening the options for experimentation beyond a hosted video-generation service.

The announcement brought generative AI into an established consumer assistant, with the rollout still to come.

The release showed that model development followed more than one approach: larger pre-training runs continued alongside explicit reasoning methods.

The change concerned how users reached answers and checked their sources by voice.

The target was maintenance across an existing codebase, where the scope and consistency of changes matter more than generating new code.

The product placed agents inside existing business processes, with integration and oversight built in.

The demonstrations, compared by some to a “DeepSeek moment,” presented an assistant that could work through a goal across several steps.

The work supported training large models across less tightly connected hardware.

The submission showed how an AI developer sought to shape the conditions under which its industry would operate; the plan itself followed in July.

The release combined a new model offering with free access for individual users.

The experiment brought the question of AI in judicial decisions into a mainstream newspaper.

The tools addressed different parts of building robotics and autonomous systems: generating training material, adapting it and reasoning about physical scenes.

The release extended the reasoning-model approach beyond text, with published evaluations on visual and mathematical tests.

The measure expressed task difficulty in human working time, giving agent progress a unit that could be tracked across models.

The agreement expanded SoftBank’s planned investment in the hardware supporting AI.

The release preserved for study how a landmark neural network was built.

The release supplied components for voice applications, converting speech into text and generating spoken responses.

The release continued development of a general-purpose model alongside the company’s R1 reasoning family.

The benchmark sought to expose gaps that earlier tests no longer captured as clearly.

The release brought a new generation of reasoning models into competition, under an experimental label.

The capability connected visual production to an ongoing dialogue, in which an image could be refined step by step.

The combination brought an AI developer and a major distribution platform under one corporate arrangement.