Elon Musk rarely admits failure, but this week he took to X (formerly Twitter) to publicly acknowledge that the current public-facing Grok 4.2 model has serious shortcomings. He also revealed how the internal version numbering maps to public releases: Grok 4.2 is based on internal version 8, a 0.5T-parameter model trained on Hopper architecture GPUs, with what he called "major defects" in training data quality, comprehensiveness, and proportionality.
https://twitter.com/elonmusk/status/2055298325994164377
Meanwhile, internal version 9 has completed training. It packs 1.5T parameters, is optimized for Blackwell architecture GPUs, and represents a complete overhaul in data processing, training recipe, and scale. Musk said version 8 and 9 "are not even in the same league." He also mentioned that data from Cursor — the AI coding tool acquired by SpaceX — will be incorporated in subsequent training, the first time integration at the data level has been discussed since SpaceX acquired Cursor.
In the same week, The Information reported that SpaceXAI's pre-training team has been reduced to just a handful of people after its head departed, with over 50 core researchers leaving since the acquisition. Musk had previously promised a new base model every two weeks. The planned 1T-parameter Grok 4.4 was supposed to ship in early May but hasn't appeared. Now a 1.5T version has been revealed that matches the specs originally slated for Grok 4.5 — leaving it unclear whether 4.4 has been skipped entirely.