Grok 4

O xAI

Založeno 2023 · Palo Alto, USA

xAI was founded in March 2023 by Elon Musk together with researchers from DeepMind, OpenAI, Google Research, Microsoft Research and Tesla including Igor Babuschkin and Tony Wu. The company's stated mission is to build maximally truth-seeking AI. Its flagship product is Grok, an LLM-based assistant deeply integrated into X (formerly Twitter) and available via a standalone app and API. Grok 1 weights were released under Apache 2.0 in November 2023 (314B-parameter MoE), Grok 1.5 added 128K context and vision in early 2024, Grok 2 shipped in mid-2024, Grok 3 launched in February 2025 trained on the Colossus supercluster, and Grok 4 followed in mid-2025 as the next frontier flagship with enhanced reasoning. xAI raised over $12 billion in 2024 across two financing rounds with backers including Andreessen Horowitz, Sequoia, Valor Equity, Lightspeed and Saudi funds. Beyond Grok the company operates the Colossus cluster in Memphis, regarded as the world's largest dedicated AI training site at launch.

Architektura

Decoder-only Transformer (frontier reasoning successor to Grok 3)

Grok 4 is xAI's successor to Grok 3, released in mid-2025 as the lab's frontier reasoning model. It was trained on the expanded Colossus supercluster in Memphis, which grew from roughly 100,000 NVIDIA H100 GPUs at Grok 3 launch to a target of 200,000+ GPUs by Grok 4. xAI describes the training compute as approximately 10x that of Grok 3. The model is a decoder-only Transformer pretrained on a multi-trillion-token mixture of web text, code, scientific papers, math content and the continuously growing feed of public posts on the X platform. Post-training relies heavily on large-scale reinforcement learning against verifiable rewards, particularly for math, coding and tool-use trajectories, building on the Think and high-effort reasoning modes introduced in Grok 3. Grok 4 ships with an upgraded DeepSearch agent that performs multi-step web and X-platform retrieval before reasoning over the gathered evidence, and supports native tool use, function calling and image input. Context window is increased to 256K tokens. xAI positions Grok 4 as competitive with OpenAI o3 and Gemini 2.5 Pro on math and science benchmarks, with the differentiator being real-time access to X data and a lower refusal rate.

Parametry

Undisclosed (estimated multi-hundred billion parameters)

Kontext

256 000 tokenů

Schopnosti

• 256K token context window
• Reinforcement-learning-trained reasoning modes with adjustable thinking budget
• Upgraded DeepSearch agent for X + web retrieval
• Trained on ~200K H100 Colossus cluster
• Strong AIME, GPQA and ARC-AGI scores at launch
• Vision input for images and screenshots
• Function calling and structured JSON output
• Native integration with X timeline and DMs in the X app
• Available via xAI API, X Premium+ and the Grok mobile app
• Real-time fresh data through built-in retrieval
• Best for: agentic research, real-time question answering, math/code reasoning, X-platform assistants.

Trénování a licence

Multi-trillion-token pretraining mixture of web text, code, math, scientific papers and X public posts. Knowledge cutoff updated continuously through retrieval. Post-training applies RLHF and large-scale reinforcement learning against verifiable rewards on math, code and tool-use tasks.

Licence: Proprietary commercial license via xAI API, X Premium+ and the Grok app. Weights not publicly released for Grok 2, 3 or 4.

Bezpečnostní testy: xAI publishes a risk-evaluation summary on x.ai. The model is intentionally less restrictive than competitor frontier models; external evaluations are limited.