xAI introduced Grok-1, pairing a frontier language model trained on a custom JAX and Rust distributed stack with live retrieval over real-time X posts.
Grok
Also known as: Grok-1 · Grok-2 · Grok 3 · Grok 4 · Grok 4.7
Plain English
xAI's flagship family of frontier language, reasoning, and multimodal models trained on the Memphis Colossus GPU superclusters and integrated across the xAI API, X, Grok Build, and Grok Bot.
Technical Overview
Developed by xAI, the Grok lineage began with the open-weight 314B MoE Grok-1 checkpoint before scaling pretraining and reinforcement learning across 100K–200K+ GPUs on Colossus in Grok 3 and Grok 4/4.7. Recent generations unify controllable test-time reasoning, native tool use, real-time web and X search, and terminal/team agent harnesses.
Timeline (7)
2023
Release2024
ReleasexAI released the base checkpoint of Grok-1 (314B parameters, 8-expert Mixture-of-Experts with top-2 routing) under Apache 2.0, followed on March 28 by Grok-1.5 expanding the context window to 128K tokens.
- Release
xAI launched Grok-2 and Grok-2 mini, delivering major gains on coding, math, and vision-language benchmarks and powering the commercial
api.x.aideveloper platform. 2025
ReleaseTrained on xAI's 200,000-GPU Colossus supercluster, Grok 3 and Grok 3 mini introduced large-scale reinforcement learning for extended test-time reasoning alongside the DeepSearch multi-step synthesis agent.
- Release
xAI released Grok 4 and Grok 4 Heavy, training the model to invoke code interpreters and search tools natively within its chain-of-thought and introducing parallel multi-agent test-time consensus in the Heavy tier.
2026
ReleasexAI rolled out Grok Imagine 1.5 (
grok-imagine-video-1.5-preview) in the xAI API, extending the Grok family to 720p image-to-video and text-to-video synthesis alongside Video Generation Models (Sora, Veo, Kling, Wan).- Release
xAI released Grok 4.7, optimizing long-horizon tool calling, repository-scale software engineering in Grok Build, and multi-user work routines in Grok Bot.
See also (14)
Open full graphGrok
9 more connections
- GPTOutgoing · Alternative to
- Mixture of Experts (MoE)Outgoing · Implements
- Reasoning & Test-Time ComputeOutgoing · Implements
- Agent HarnessOutgoing · Mentions
- Context Window & RoPEOutgoing · Mentions
- Multimodal & Vision-Language ModelsOutgoing · Mentions
- Post-Training (SFT, RLHF, DPO & GRPO)Outgoing · Mentions
- Tool Use & Function CallingOutgoing · Mentions
- Video Generation Models (Sora, Veo, Kling, Wan)Outgoing · Mentions