Grok 3 Mini
Version: 1
Grok 3 Mini is xAI's lightweight reasoning model, trained to solve agentic, coding, mathematical, and deep science problems with test-time compute. The raw reasoning traces will be returned to the user for inspection, with "low" and "high" thinking budgets to adjust how long the model thinks.
Grok 3 Mini supports a 131,072 token context window for understanding codebases and enterprise documents. The model was trained via reinforcement learning with a focus on reasoning for agentic coding tasks, and excels at utilizing tools to solve complex logical problems in novel environments.
Model developer: xAI
Supported languages: English, Spanish, French, Afrikaans, Arabic, Bengali, Welsh, German, Greek, Indonesian, Icelandic, Italian, Japanese, Korean, Latvian, Marathi, Nepali, Punjabi, Polish, Russian, Swahili, Telugu, Thai, Turkish, Ukrainian, Urdu, and Chinese.
Model Release Date: May 19, 2025
Intended Use
Primary Use Cases
Grok 3 Mini is a lightweight yet powerful thinking model delivering intelligence with cost-efficiency. The model is optimized for logic-based tasks, such as:- Coding environments: working inside codebases and local development environments.
- Agentic workflows: building robust LLM ontologies and agent architectures.
- Reasoning tasks: difficult mathematics and science based questions.
Core Capabilities
- Extended Context Length: With an extended context length of up to 16k tokens (131K coming soon), Grok 3 Mini processes and understands vast datasets in a single pass—ideal for comprehensive analysis of large documents or complex workflows.
- Exposed Reasoning Tokens: Unlike traditional black-box thinking models, Grok 3 Mini offers unparalleled transparency, letting its users inspect its reasoning tokens. This transparency is a game-changer for enterprises and educators needing to understand the “why” behind answers—reflecting xAI’s commitment to openness.
- Steerability & Chain of Command: Grok 3 Mini is extremely steerable and follows instructions closely. The model is less likely to refuse queries, providing more helpful responses while maintaining safety and ethical standards.
- Reasoning effort parameter: For more fine grained control over the model’s performance, Grok 3 Mini supports the reasoning effort parameter, which allows users to adjust the model’s thinking effort with options for low and high reasoning levels.
- Structured outputs: Grok 3 Mini model supports structured outputs, enabling developers to specify JSON schemas for AI-powered automations.
- Functions and Tools support: Similar to other xAI models, Grok 3 Mini supports functions and external tools that enable enterprises to build agentic workflows.
Grok 3 Mini (High) Benchmark Performance Overview
To understand its capabilities, xAI evaluated Grok 3 Mini (High) on a variety of benchmarks using their internal benchmarking platform. Grok 3 Mini (High) delivers state-of-the-art results across diverse academic benchmarks among non-reasoning models, including: Graduate-level science knowledge (GPQA), General knowledge (MMLU-Pro), and Math competition problems (AIME). Below is a high-level overview of the model quality on representative benchmarks:| Category | Benchmark | Grok 3 Mini (High) Score (%) |
|---|---|---|
| Math Competition | AIME 2024 | 90.7 |
| Graduate-Level Reasoning | GPQA | 80.3 |
| Code Generation | LiveCodeBench | 74.8 |
| Multi-Task Language Understanding | MMLU-pro | 82.8 |
| Average | 82.2 |
Model Specifications
Context Length131072
LicenseCustom
Last UpdatedMay 2025
Input TypeText
Output TypeText
PublisherxAI
Languages27 Languages