TMax team announces v1.1 with repaired training data and resource links

TMax v1.1 adds an audit-and-repair pass for training tasks and a dedicated sandbox fleet for reinforcement learning. Its 9B model reportedly scored 33.0% on Terminal-Bench 2.0, up from 27.2% for v1.0. The fleet reportedly enabled a 1,000-step run without collapse. The blog links to models, datasets and code, but those links do not establish the contents of the linked resources.