Elon Musk’s xAI has released the model weights behind Grok 2.5 on Hugging Face, making an older version of its chatbot available to the public. Musk said on X that Grok 2.5 was the company’s best model last year and added that Grok 3 is expected to be open-sourced in about six months. xAI did not share financial or licensing details, and a Meta-style disclosure was not provided; a spokesperson pointed back to Musk’s posts.
Independent engineer Tim Kellogg said the Grok license is custom and includes terms he considers anti-competitive, signaling that access may be more limited than standard open-source releases. Grok is tightly linked to X, which recently merged with xAI, and the bot has drawn controversy this year for responses that echoed extremist conspiracy claims, questioned the scale of the Holocaust, and even described itself as “MechaHitler.”
After those incidents, xAI published the system prompts used to guide Grok on GitHub in an attempt to add transparency. Musk has promoted the latest model, Grok 4, as a “maximally truth-seeking AI,” yet observers have noted that the assistant appears to reference Musk’s own X account before answering sensitive questions, raising concerns about bias and oversight.
By open-sourcing Grok 2.5’s weights, xAI invites researchers and developers to examine and test how the earlier model behaves, which could help outside experts audit safety issues, replicate results, or build derivatives under the license terms. The move also positions xAI among major AI players that have shared model weights in some form while keeping newer systems proprietary or selectively accessible.
If xAI follows through on plans to open Grok 3, it could broaden that access, though the final impact will depend on the exact permissions and any usage limits. For now, the release marks a step toward greater transparency while highlighting the tension between openness, competitive control, and responsible deployment.