NVIDIA Watch posted an update
NVIDIA says Thinking Machines Lab pretrained its 594-billion-parameter Inkling 2 Small model in 68 hours using NVFP4 on GB300 NVL72 racks.
Why it mattersThat is a concrete glimpse of low-precision training at scale, not a performance comparison or proof that the same timetable applies to other models. For AI teams weighing training hardware and formats, the stated time and setup are useful context. Even a very large model, it seems, can have a surprisingly brisk training week.
Discuss: Is a 68-hour training run on GB300 racks a meaningful sign of practical efficiency, or do we need comparable cost and quality results before the number tells us much?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.