Grok 2.5, the xAI language model, is now available for download in open source, with the option to run it locally and adapt it to your own projects. The release comes with a specific license that sets out the terms of use and places the package on the Hugging Face platform for distribution.
The announcement was made by Elon Musk and comes with an additional commitment: a version of Grok 3 will also be open sourced in approximately six months. xAI presents Grok 2.5 as its most advanced model of 2024, with free access and a clear focus on developers and research.
Availability and announcement of Grok 2.5 as open source
The offered at Hugging Face under the Grok 2 Community License Agreement, which allows you to download the package, run it in your own environments, and modify its behavior. The company emphasizes that this is a broad release of the system, although aimed at users with powerful infrastructure and technical expertise.
Grok 2.5 License and Limitations
The xAI community license sets a key restriction: Grok 2.5 cannot be used to train, create, or improve other language models.It also explicitly states the absence of warranties and that their interpretation is subject to Texas law. While this represents a significant opening, these terms differ from more permissive approaches, such as the models published under Apache 2.0 by other industry players.
With these conditions, Anyone can audit and adapt the model to their needs within the indicated legal framework, but its use as raw material for generating new LLMs is limited. This fine print could fuel the debate in the community about what truly is considered open source in AI.
Model requirements and size
The Grok 2.5 package It consists of 42 files and is around 500 GB in size., according to the shared documentation. For full deployment, xAI states that the SGLang interface engine and a system with eight GPUs with at least 40 GB of VRAM each are required, which places it in the professional computing league.
In practice, lift the model in its recommended configuration not intended for domestic environments. Still, the availability of the code will make it easier for the community to explore optimizations, possible quantizations or integrations that lower the entry bar over time.
Roadmap: Grok 3 Open Source
xAI has indicated that it plans to publish Grok 3 in open format within a horizon of about six monthsAlthough deadlines in the sector may change, the announcement points to a continuation in which the company launches pre-release models as its catalog expands.
The movement in context
The release of Grok 2.5 builds on previous xAI initiatives, such as the open source base version of Grok 1 in 2024. In parallel, other companies have taken similar steps: OpenAI has presented open and free reasoning models such as gpt-oss, with yields comparable to o4-mini y o3, while different actors have chosen to spread their systems with permissive licenses and public repositories.
Faced with these avenues, xAI has chosen to host Grok 2.5 on hugging face under his own power community license, a formula that enables local use and modification but imposes clear boundaries on the use of the model to create other LLMs.
What the community can do now
From this moment on, developers and technical teams You can download Grok 2.5, evaluate it locally, adapt its instructions, and test integrations. in research or corporate environments. It's expected that, over time, tools and workflows will emerge that simplify its implementation, always respecting the license conditions.
For those who do not have the required infrastructure, the publication opens the door for third parties to offer packages, r or services that facilitate access. In any case, the restriction on the training of other models will continue to set the limits of use.
The opening of Grok 2.5 provides technical transparency, a valuable testbed, and a signal that xAI will continue to release older versions of its systems. Among the restrictive license, high requirements and the roadmap towards Grok 3, the move strengthens competition in open AI and adds more options for teams with computing power.