Elon Musk, founder and CEO of xAI, announced in a post on X that the company’s Grok 2.5 model from last year is now available as an open-source model. The model’s weights can be downloaded from Hugging Face.
This reflects a pattern the company has adopted, making some of its models available as open source, similar to what they did with the first version of Grok last year.
Musk also said that the company’s Grok 3 model will be made open source in 6 months.
Currently, Grok-3 offers comparable performance with models like Anthropic’s Claude 4 Sonnet, GPT 4.1, whereas the company’s latest Grok 4 model leads several benchmarks, even outperforming OpenAI’s GPT-5 in certain scenarios.
Source: Artificial Analysis
Currently, Grok 3 is the available model for all free users on the platform, and the company has also recently announced limited free access to its latest Grok 4 model.
Sebastian Raschka, an AI researcher, said in response to Musk’s post on X, “It’s quite cool that the open source community gets the real, full-sized model that was used in production (as opposed to the lite or spin-off versions).”
Along with xAI, Google, and recently OpenAI, have also announced several open-source models. Google has been maintaining and updating the Gemma family of open models for quite some time, and although they do not match the capabilities of the Gemini family of models, the company has released several purpose-built Gemma models.
Recently, Google announced Gemma 3 270M, a compact 270-million parameter model intended for task-specific fine-tuning and efficient on-device deployment.
Similarly, OpenAI released two new open-weight language models, gpt-oss-120b and gpt-oss-20b—under the Apache 2.0 license.
Meanwhile, Chinese AI labs have also been dominant when it comes to releasing high-performance open source models. The recent announcements of Ziphu AI’s GLM 4.5 models and Moonshot AI’s Kimi K2 models were observed to offer some of the best performance by open models on benchmark evaluations, following the footsteps of DeepSeek, which once caused quite a storm in the ecosystem.
As the industry awaits the next iteration of the DeepSeek reasoning model and a successor to the R1, a new hybrid model called DeepSeek-V3.1 was released a few days ago.
