xAI today published the weights for Grok-1, the 314-billion-parameter mixture-of-experts model, under an Apache 2.0 license on GitHub. The release includes JAX code for loading and inference, along with a torrent link for the checkpoint.

The model uses eight experts with two active per token, a 64-layer architecture, and a context window of 8,192 tokens. However, the release is limited to raw base weights without instruction tuning or training data. The repository’s README explicitly states the implementation ‘is not efficient’ and requires ‘a machine with enough GPU memory.’

On Hacker News, where the announcement drew 1,170 points and 419 comments, the reception was mixed. Several commenters noted that while the parameter count makes Grok-1 the largest open-weight model yet, smaller models like Mixtral 8x7B reportedly outperform it on many metrics. The debate quickly turned to semantics: users argued whether ‘open weights’ qualifies as open source, with one commenter calling it ‘like calling a binary open source because you can download it.’ Others pointed out that without training data, fine-tuning risks catastrophic forgetting.

For now, the weights are out, but the training data remains behind closed doors.