enkukliosAn encyclopedia of AISearch

5 min read

Frontier AI Labs Around the World

On this page

This article surveys the organizations building the most capable AI systems—large language models, multimodal systems, agents, and world models—across the US, Europe, China, and other regions. It complements Frontier LLMs, World Models, and Modern Model Variations, and links back to the Frontier AI hub.

United States

  • OpenAI (San Francisco) — ChatGPT, the GPT series (GPT-4, GPT-5.x, GPT-6 Astra and GPT-6 Sol), the o1/o3 reasoning line, Codex coding agents, GPT Image, and (until 2026) the Sora video model. OpenAI’s most capable public release is the GPT-5.6 Sol class, and GPT-6 Astra became broadly available in September 2026 before a delayed GPT-6.1. It also publishes the GPT Image models and the Codex agent. openai.com
  • Anthropic (San Francisco) — the Claude series, including Claude 3/3.5, Claude 4 Opus and Sonnet, Claude Opus 4.6 and Sonnet 4.6, the security-oriented Claude Mythos 5, and the Opus 5.5 / Sonnet 5.5 releases of late 2026. Claude ships with agent harnesses: Claude Code, Cowork, Design, and Science. Safety research (mechanistic interpretability, alignment) is a defining trait. anthropic.com
  • Google DeepMind — the Gemini family (Gemini 1.x, 2.x, and the 3.x generations, including Gemini 3.5, 3.6 Flash, 3.7 Flash, and 3.8 Flash), open Gemma models, AlphaFold for biology, and world models (Genie 3, GenCast). Gemini 2.5 Pro was Google’s first explicitly “thinking” model; the 3.x line added agentic and multimodal “Flash” variants. deepmind.google
  • Meta AI / Meta Superintelligence Labs — the open Llama family (Llama 2, Llama 3, Llama 4). In April 2026 Meta Superintelligence Labs released Muse Spark as a replacement for Llama in Meta’s chatbots, with the Muse Spark 1.1 and Muse Glimmer variants, plus open-weight world/video models (V-JEPA 2). ai.meta.com
  • xAI / SpaceXAI — the Grok models, integrated with X and Tesla’s Optimus. Grok-1 was open-sourced in 2024; Grok 2, Grok 3, Grok 4, Grok 4 Heavy, Grok 4 Fast, and Grok 4.5 followed, with Grok 4.5 co-developed with the Cursor subsidiary. x.ai
  • Microsoft and Amazon — partner with and host frontier models (OpenAI and Anthropic on Azure/Bedrock) rather than developing their own leading base models.

Europe

  • Mistral AI (France) — sovereign open models. Mistral Large 3 (December 2025) is a 675-billion-parameter mixture-of-experts model with 41 billion active parameters; Mistral Large 4 (1.05T parameters) entered preview in October 2026. Its chatbot, Le Chat, was rebranded Mistral Vibe in 2026. Mistral is central to the EU’s digital-sovereignty push. mistral.ai
  • Aleph Alpha (Germany) — Luminous models and enterprise AI. aleph-alpha.com
  • DeepL (Germany) — neural translation systems.
  • Other EU labs focus on open, sovereign, and regulated AI, often funded through EU and national programs.

China

  • DeepSeek (Hangzhou) — open-weight models: DeepSeek-V3.2 (December 2025, 685B parameters, MIT license) with a custom DeepSeek Sparse Attention mechanism; the V4 series preview (April 2026) with DeepSeek-V4-Flash (284B parameters, 13B active) and DeepSeek-V4-Pro (1.6T parameters, 49B active); and the R1 reasoning line (DeepSeek-R1, January 2025; R1-0528). Models use one-million-token contexts and are MIT-licensed. deepseek.com
  • Alibaba Cloud (Qwen) — Qwen open models. The latest, Qwen3.8 (July 2026), is a 2.4-trillion-parameter model with 95 billion active parameters, the second-largest open-weight model after Kimi K3; a distilled Qwen3.8 27B is Apache-licensed. Earlier generations include Qwen3, Qwen3.6-35B-A3B, Qwen-Audio, Qwen2-Math, and the QwQ reasoning line. A local Qwen powers Apple Intelligence in China. qwenlm.github.io
  • Z.ai (formerly Zhipu AI) (Beijing) — the GLM family. GLM-5 (February 2026, 744B parameters) through GLM-5.3 (August 2026, 744B, 40B active) are MIT/Apache-licensed open-weight models; GLM-5.3-Flash followed as an open release. Z.ai is China’s first major LLM company to list publicly (Hong Kong, January 2026) and was placed on the US Entity List in 2025. z.ai
  • Moonshot AI (Beijing) — the Kimi series of long-context, agentic open-weight models: Kimi K1.5, K2, K2.5 (multimodal, January 2026), K2.6, K2.7 Code, and Kimi K3 (July 2026, 2.8T parameters, 104B active)—the largest open-weight model ever released. Its “Swarm” feature chains hundreds of sub-agents for deep research; the Mooncake platform serves 100 billion tokens daily. moonshot.ai
  • ByteDance — the Doubao chatbot (330 million users by May 2026) powered by the Doubao Large Model on Volcano Engine; the Seed proprietary and Seed-oss open models; Seedream text-to-image; and Seedance 2.0 image/text-to-video. OmniHuman-1 creates videos from a single image. doubao.com
  • Baidu — ERNIE 4.5 (foundation model) and ERNIE X1 (reasoning model), released March 2025; Ernie Bot chatbot; the Apollo autonomous-driving stack. baidu.com
  • Huawei — the Pangu models and Ascend hardware. iFlytek — the Spark models.
  • China’s labs emphasize open weights, scale, state-supported research, and (for DeepSeek, Z.ai, Moonshot) increasingly open, MIT/Apache-licensed releases that rival US frontier models.

Other Regions

  • Korea — LG AI Research leads a consortium developing the sovereign open model K-EXAONE 2.0 (750B parameters, 37B active, July 2026), part of the government’s “AI Foundation Model” project; Samsung has its Gauss on-device models. ESTsoft participates in the K-EXAONE consortium.
  • Japan — NTT and academic initiatives, plus partnerships (e.g., ESTsoft’s Perso AI with NTT and Nihon Kotsu for taxi AI).
  • India and the UAE — national AI programs and sovereign-model efforts; smaller labs focus on local-language models and deployment.

World Models and Other Variations

Frontier labs increasingly build world models alongside LLMs: Google DeepMind’s Genie 3, Waymo’s Waymo World Model, Nvidia’s Cosmos 3, Alibaba’s Happy Oyster, Meta’s V-JEPA 2, and LeCun’s LeWorldModel. These are covered in depth in World Models. Architectural variations such as mixture-of-experts, mixture-of-transformers, reasoning models, and multimodal systems are covered in Modern Model Variations.

Why It Matters

Frontier AI is a global field. Understanding labs outside the US—especially China’s open-weight DeepSeek, Qwen, GLM, and Kimi families—is essential to map capability, openness, and governance. Open-weight releases now rival closed US frontier models, and world-model research is moving beyond text.

Sources

Back to Recent Topics · See Frontier AI hub · Sub-article index.

Related articles

History of Artificial Intelligence Recent Topics in Artificial Intelligence