Today we released Grok-4 weights under an Apache 2.0 license. The model is competitive with the frontier on reasoning and coding, behind on long-context recall, and ahead on Chinese and Spanish. Here's the thinking.

Open weights are good for the field

Closed labs argue that releasing weights is dangerous. We disagree. Most of the safety work that matters happens at the application layer — RLHF and constitutional alignment harden a model against specific abuse patterns, but they don't make a base model safe in isolation. Releasing weights lets the broader research community reproduce, audit, and improve our safety techniques.

What's in the release

• 7B, 70B, and 400B parameter base models
• Instruction-tuned variants of all three
• The training data card (without the actual data)
• The eval suite we used internally
• A small dataset of red-team prompts and our refusal heuristics

What's not

We're not releasing the multimodal vision tower yet. The reason is boring — it's still being trained. Expect a separate vision release in Q2 once we've finished the safety evaluations specific to images.

Try it

The weights are on Hugging Face. The inference code is on GitHub. If you find weaknesses or improvements, file an issue or send a PR. We're committing to a 2-week response SLA on security-relevant reports.

We're also publishing a paper on the training methodology next month. The short version: bigger isn't always better, and the data mixture matters more than the model architecture beyond a certain scale. The paper has the details.

Happy hacking.