Grok 3 and Grok 3 mini
2025-02-19 · Publisher : xAI · Domain : Artificial Intelligence
xAI introduces Grok 3, its flagship model trained on the Colossus supercomputer, and Grok 3 mini, a cost-efficient reasoning variant, with the Think and DeepSearch modes.
What changes
- Grok 3 is presented as trained with ten times more compute than previous state-of-the-art models, according to xAI.
- Two reasoning models in beta: Grok 3 (Think) and Grok 3 mini (Think), trained with reinforcement learning, with visible reasoning that can last from a few seconds to several minutes.
- DeepSearch, xAI's first agent, combines reasoning and Internet search.
- Grok 3 mini targets STEM tasks that require less general knowledge.
### Specifications - Announced context window: 1 million tokens (eight times more than xAI's previous models). The API limit at release was 131,072 tokens according to a third-party aggregator. - Modalities: text and image input (MMMU and EgoSchema published). - Availability at launch: Grok.com and X for Premium and Premium+ subscribers (Think and DeepSearch first for Premium+). API announced “in the coming weeks”. - Knowledge cutoff: not disclosed in the announcement. ### Pricing Pricing not disclosed at launch (see the “Grok 3 in the xAI API” entry for the API). ### Benchmarks Scores published by the vendor. | Benchmark | Grok 3 (Think) | Grok 3 mini (Think) | |---|---|---| | AIME 2025 (cons@64) | 93.3% | not disclosed here | | AIME 2024 | not disclosed here | 95.8% | | GPQA Diamond | 84.6% | not disclosed here | | LiveCodeBench | 79.4% | 80.4% | | Chatbot Arena (Elo) | 1402 | not disclosed | ### Key takeaways First xAI generation oriented toward reasoning, with an integrated search agent. The models were still being trained at launch, and API access only came afterward.
Links
Tags : LLM, xAI, Grok, Grok 3, Raisonnement