⏳ In Brief
- xAI released Grok 2.5 weights on Hugging Face under a custom community license.
- Download totals about 500 GB, with 42 files, and SGLang inference instructions.
- Running locally requires eight GPUs with over 40 GB each, and tensor parallelism.
- The license allows commercial use with guardrails and bans training other foundation models.
- Elon Musk says Grok 3 could be open-sourced in about six months.
Breaking: xAI releases Grok-2.5 weights under a custom community license
xAI has published the Grok 2.5 model weights on Hugging Face, with instructions for local inference. The release centres on downloadable checkpoints, not a training pipeline, and ships under the Grok 2 Community License rather than an OSI licence.
xAI listed Grok 2.5 for download on Hugging Face, pointing to SGLang for serving and citing a total size of nearly 500 GB across 42 files.
The @xAI Grok 2.5 model, which was our best model last year, is now open source.
Grok 3 will be made open source in about 6 months. https://t.co/TXM0wyJKOh
— Elon Musk (@elonmusk) August 23, 2025
What you can actually download and run
The model card provides weights, a tokenizer path, and commands to serve with SGLang using FP8. It specifies tensor parallelism of 8, reflecting multi-GPU assumptions for throughput and memory. A chat template is required for correct formatting.
Hardware notes state you need eight GPUs, each with more than 40 GB of memory, to host the checkpoint as documented. The download is large, about 500 GB, and retries may be needed to complete the transfer.
What is included today
- Weights and tokenizer references
- SGLang inference commands and chat template
- A community license file on the repo
The licence, in plain terms
The Grok 2 Community License permits non-commercial use, and commercial use if you follow xAI’s Acceptable Use Policy. It prohibits using the model or outputs to train other foundation models, except for allowed Grok 2 fine-tuning.
Distribution requires including the licence and displaying “Powered by xAI.” The licence is revocable on breach, sets Texas governing law, and clarifies that xAI retains rights in the materials while outputs are generally unrestricted.
“The xAI Grok 2.5 model, which was our best model last year, is now open source… Grok 3 will be made open source in about 6 months.”
Open weights versus open source, and the community reaction
This is open-weights, not open source in the OSI sense. The release includes checkpoints and inference notes, not the complete training data or end-to-end pipeline. That matters for reproducibility and academic audits.
Early community threads highlight the custom licence, noting restrictions and the revocation clause. Some describe it as “hostile” compared with permissive licences, while others welcome any release of large weights.
Why this matters for builders and researchers
Having production-used weights enables behavioural study, jailbreak testing, and safety evaluations offline. It supports research into prompt formatting, decoding strategies, and alignment choices without per-token API costs.
Practical deployment is still compute-heavy. The recommended setup expects eight GPUs and advanced inference tooling, which limits accessibility to well-resourced labs and infrastructure teams.
The release benefits evaluation and reproducibility, but the hardware bar and licence terms mean it is not a drop-in for every developer.
What remains unconfirmed or evolving
The model card does not list a clear parameter count for Grok 2.5. Some coverage infers prior MoE scales, yet the repository itself avoids pinning a public size. Treat external estimates as unconfirmed.
Details on training data, safety filters, and evaluation suites are not bundled with the weights. Those omissions limit end-to-end replication, though they do not prevent meaningful empirical testing on behaviour.
Conclusion
xAI’s Grok 2.5 is now downloadable, with weights and inference guidance available publicly. The package is useful for study and prototyping, while the license places clear boundaries on commercial and derivative uses.
For teams with adequate hardware, the release offers a chance to examine a frontier-class checkpoint outside an API. Watch for any clarifications on scope, plus whether a future Grok 3 release follows this template.
📈 Trending News
25th August 2025
- Siri reboot on the table, Apple in talks to use Gemini
- South Park skewers AI sycophancy, turning ChatGPT into a yes-machine
- Adobe launches Acrobat Studio to turn PDFs into AI-powered workspaces
- Meta nabs sixth Apple AI leader in seven weeks
- Anthropic aims to raise up to $10B amid strong investor interest
For more AI stories, visit AI News on our site.
