xAI open-sources Grok-1's model weights β a 314-billion-parameter MoE model
xAI has open-sourced the model weights and network architecture of βGrok-1,β the foundation model behind the βGrokβ chatbot, under the Apache 2.0 license. Releasing such a large-scale 314-billion-parameter Mixture-of-Experts model under a permissive license that allows commercial use drew significant attention.
Details
- Model scale: A 314-billion-parameter Mixture-of-Experts model designed so that roughly 25% of the total weights are activated per token
- Training infrastructure: Built using a custom training stack based on JAX and Rust
- Scope of release: A raw checkpoint from the pretraining phase completed in October 2023 β an unadjusted base model that has not been fine-tuned for chat or other applications
- License: Released under the Apache 2.0 license, allowing free modification and redistribution, including for commercial use
- Distribution: Implementation guides and related resources were published on GitHub (xai-org/grok)
What happened next
Open-sourcing Grok-1 signaled that xAI cared not just about its own commercial service but also about contributing to the research community. Shortly after this release, in late March 2024, βGrok-1.5,β with improved conversational performance, was announced, and xAIβs pace of model development accelerated further.
Source https://x.ai/news/grok-os