Overview
Grok 2 is xAI's flagship large language model from 2024, released in beta on August 13 and broadly to X Premium and Premium+ subscribers around August 20, 2024. It shipped alongside a lighter Grok 2 mini and was xAI's first model to add vision and, on X, image generation (initially via Black Forest Labs' FLUX.1). An early checkpoint tested anonymously as "sus-column-r" on the LMSYS Chatbot Arena leaderboard, where xAI said it outranked Claude 3.5 Sonnet and GPT-4-Turbo by Elo at the time.
On xAI's published academic benchmarks, Grok 2 scored 87.5% on MMLU, 75.5% on MMLU-Pro, 56.0% on GPQA, 76.1% on MATH and 88.4% on HumanEval, plus strong multimodal results of 66.1% on MMMU, 69.0% on MathVista and 93.6% on DocVQA. In December 2024 xAI brought the line to its enterprise API as grok-2-1212 (and grok-2-vision-1212) with a 131,072-token context window, priced at $2 per million input tokens and $10 per million output tokens.
Grok 2 is a sparse mixture-of-experts transformer: roughly 270 billion total parameters across 64 blocks with 8 experts, of which 2 activate per token, for about 115 billion active parameters. In August 2025 xAI open-sourced the Grok 2.5 weights on Hugging Face under the source-available Grok 2 Community License Agreement (which forbids using the model to train competing models). Grok 2 is now a legacy model, superseded by Grok 3 and the Grok 4 family.
| Released | 2024-08-20 |
|---|---|
| License | Grok 2 Community License Agreement (source-available) |
| Weights | Open weights |
| Parameters | 270B total · 115B active (mixture-of-experts) |
| Context | 131K |
| Architecture | Mixture-of-Experts transformer — 64 blocks, 8 experts with 2 active per token (270B total, ~115B activated) |
| Modalities | Text, Vision |
| Status | Legacy; weights open-sourced as Grok 2.5 in August 2025 |
Benchmarks
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $2.00 per million tokens |
|---|---|
| Output | $10.00 per million tokens |
grok-2-1212 / grok-2-vision-1212 enterprise API pricing, announced December 2024.
Strengths
- Strong 2024-era general knowledge — 87.5% MMLU and 75.5% MMLU-Pro (state-of-the-art on MMLU-Pro at release)
- Capable coding for its generation — 88.4% on HumanEval
- Excellent document and visual reasoning — 93.6% DocVQA and 69.0% MathVista
- Mixture-of-experts design activates only ~115B of 270B parameters per token
- Weights later open-sourced (as Grok 2.5) under a source-available license for self-hosting
Best for
- General-purpose chat and reasoning assistants (Grok on X)
- Document understanding and visual question answering
- Coding help and code generation
- Multilingual instruction-following via the grok-2-1212 API
- Self-hosted experimentation with the open-sourced Grok 2.5 weights
How to access
| Provider | Model ID |
|---|---|
| xAI API ↗ | grok-2-1212 |
| OpenRouter ↗ | x-ai/grok-2-1212 |
Grok (flagship) — every version
The full lineage of the Grok (flagship) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
FAQ
When was Grok 2 released?
xAI announced Grok 2 (and Grok 2 mini) in beta on August 13, 2024 and rolled it out to X Premium and Premium+ users around August 20, 2024. The grok-2-1212 API version followed in December 2024.
How big is Grok 2 and what is its architecture?
Grok 2 is a sparse mixture-of-experts transformer with roughly 270 billion total parameters across 64 blocks and 8 experts; only 2 experts fire per token, so about 115 billion parameters are active per forward pass.
How much does the Grok 2 API cost?
When xAI added grok-2-1212 and grok-2-vision-1212 to its enterprise API in December 2024, pricing was $2 per million input tokens and $10 per million output tokens, with a 131,072-token context window.
Are Grok 2's weights open?
Yes. In August 2025 xAI open-sourced the Grok 2.5 weights on Hugging Face under the source-available Grok 2 Community License Agreement, which permits use and modification but forbids training competing models with it.