xAI has made Grok-1 easier for outsiders to inspect, download, and potentially adapt. The company released the model’s base weights and network architecture, putting a major AI system into the open-weights category at a moment when Elon Musk is pressing OpenAI over how closed its models have become.
The move is important, but it should not be confused with a consumer-friendly chatbot launch. Grok-1 is large, raw, and not tuned for conversation, which means the release is more useful to researchers and advanced developers than to ordinary users looking for a ChatGPT-style assistant.
What xAI released
On Sunday, Elon Musk’s AI company xAI released the base model weights and network architecture for Grok-1. The files were made available through GitHub and BitTorrent, with inference code on GitHub and the weights available through a Torrent link listed there.
Grok-1 is the large language model behind Grok, an AI assistant that was announced in November. Grok is available to X Premium+ subscribers who pay $16 a month on the social media platform formerly known as Twitter.
The model itself is described as a mixture-of-experts LLM with 314 billion parameters. The source article gives GPT-3’s 175 billion parameters as a comparison point, while also noting that parameter count is only a rough indicator of model complexity and potential usefulness.
The release is not the same thing as handing everyone the full product that subscribers use inside X. xAI says this is the base model, not a model fine-tuned for a specific job.
“This is the raw base model checkpoint from the Grok-1 pre-training phase, which concluded in October 2023,”
xAI also states that the model is not fine-tuned for a particular application, including dialogue. In practical terms, it can do next-token prediction, completing a text prompt with what it estimates should come next, but that does not automatically make it a polished assistant.
Why the chatbot distinction matters
The most immediate misunderstanding around a release like this is the gap between a base model and a usable chat product. A base model can be powerful, but it has not necessarily been shaped to follow instructions, hold a conversation, or behave like the assistant people expect when they hear the name Grok.
AI researcher Simon Willison, who spoke to Ars via text message, put the point plainly:
“It’s not an instruction-tuned model,”
He added that there is still substantial work required before it can operate in a conversational setting. That work would require the right skills and compute capacity, which limits who can realistically take the released checkpoint and turn it into something chat-like.
That distinction also explains why the release can be meaningful for the AI community without being immediately useful to most people. Developers can study the architecture, experiment with inference, and potentially build from the checkpoint. But users should not expect to download Grok-1 and run a polished assistant on an everyday machine.
The hardware barrier is real
Grok-1 is not small. The weights checkpoint is 296GB, which makes the model difficult to run outside datacenter-class inference hardware. The source article notes that only hardware with enough RAM and processing power is likely to load the full model at once.
For comparison, the largest Llama 2 weights file cited in the source, a 16-bit precision 70B model, is around 140GB. Grok-1’s checkpoint is much larger than that, which changes the practical audience for the release.
At the time described in the source, Ars had not seen anyone get the model running locally. There were reports that people were working on a quantized model, which could reduce the size enough to run on consumer GPU hardware. The tradeoff is that quantization would also dramatically reduce processing capability.
Willison described the challenge this way:
“It’s hard to evaluate [Grok-1] right now because it’s so big—a [massive] torrent file, and then you need a whole rack of expensive GPUs to run it. There may well be community-produced quantized versions in the next few weeks that are a more practical size, but if it isn’t at least quality-competitive with Mixtral, it’s hard to get too excited about it.”
That makes Grok-1’s release a technical event first. The open availability matters, but actual experimentation depends on hardware access, model optimization, and the willingness of outside groups to invest the effort.
Open weights, not fully open AI
The language around this release matters. Musk initially said Grok would be released as “open source” in a tweet posted last Monday. But xAI is not calling the GitHub release open source, because that term has a specific meaning in software.
The AI industry has not settled on one clear term for model releases that include code and weights but do not include everything needed to reproduce training. Some releases also include restrictions. In this case, the source describes Grok-1 as better understood as “source available” or “open weights.”
The license is still notable. xAI released the base model weights and network architecture under the Apache 2.0 license. Willison said the Apache 2 license is one of the most interesting parts of the release, contrasting it with licenses used for models like Llama 2.
He also called Grok-1 one of the largest open-weights models released so far. That gives the release significance even if its immediate practical use is limited by size and lack of instruction tuning.
The OpenAI backdrop
The timing is inseparable from Musk’s dispute with OpenAI. Musk has criticized OpenAI for not releasing its models in an open way and has sued OpenAI and its executives, accusing them of prioritizing profits over open AI model releases.
Musk was a co-founder of OpenAI but is no longer associated with the company. The source says he regularly pushes OpenAI to release models as open source or open weights, partly because many believe the company’s name implies that kind of approach.
On March 5, OpenAI responded to Musk’s allegations by revealing old emails that appeared to suggest Musk had once accepted OpenAI’s move toward a for-profit business model through a subsidiary. OpenAI also said the “open” in its name referred to making resulting products available for everyone’s benefit, rather than committing to an open-source method.
That same day, Musk tweeted, across two tweets:
“Change your name to ClosedAI and I will drop the lawsuit.”
Five days later, he announced that Grok would be released openly. The result is a release that is both a technical milestone and a public argument about what openness should mean in advanced AI.