enkukliosAn encyclopedia of AISearch

3 min read

Frontier LLMs (Large Language Models) as of late 2026

On this page

Current frontier language models span the US, Europe, and China. This article updates Large Language Models and complements Frontier Labs, World Models, and Modern Model Variations. Dates and model names reflect the state of the field in late 2026.

United States

  • OpenAI — GPT-4, GPT-5.x (including GPT-5.6 Sol), GPT-6 Astra (September 3, 2026), and a delayed GPT-6.1; reasoning models o1 and o3; Codex coding agents and GPT Image. The GPT-6.1 release was paused over deception and authorization concerns raised by researchers. OpenAI
  • Anthropic — Claude 3 Opus/Sonnet/Haiku, Claude 3.5 Sonnet, Claude 4 Opus and Sonnet, Claude Opus 4.6 (February 2026), Sonnet 4.6, Claude Mythos 5 (security-focused), Claude Fable 5 (Mythos-class), Claude Opus 5.5 (September 22, 2026), and Sonnet 5.5 (September 28, 2026). Anthropic
  • Google — Gemini 1.x, 2.x, and the 3.x line: Gemini 3, Gemini 3 Deep Think (February 2026), Gemini 3.5 family (July 21, 2026), Gemini 3.6 Flash, Gemini 3.7 Flash (August 13, 2026), and Gemini 3.8 Flash (September 2, 2026). Open Gemma models remain available. Google DeepMind
  • Meta — Llama 3.3, Llama 4 (April 2025), then Muse Spark (April 2026) as the replacement powering Meta’s chatbots, plus Muse Spark 1.1 and Muse Glimmer (August 2026). Meta AI
  • xAI / SpaceXAI — Grok 3, Grok 4, Grok 4 Heavy, Grok 4 Fast (September 2025), and Grok 4.5 (2026), integrated with X and Optimus. x.ai

Europe

  • Mistral — Mistral 7B, Mixtral 8x7B, Mistral Large (2024), Mistral Large 2, and Mistral Large 3 (December 2025, 675B MoE, 41B active); Mistral Large 4 (1.05T MoE) entered preview in October 2026. Le Chat became Mistral Vibe in 2026. Mistral AI
  • Aleph Alpha — Luminous models. Aleph Alpha

China

  • DeepSeek — DeepSeek-V3.2 (December 2025, 685B, MIT), DeepSeek-V3.2-Speciale (reasoning), DeepSeek-V4-Flash (284B, 13B active, July 2026) and DeepSeek-V4-Pro (1.6T, 49B active, August 2026) with one-million-token contexts; reasoning models DeepSeek-R1 (January 2025) and R1-0528. DeepSeek
  • Alibaba (Qwen) — Qwen3, Qwen3.6-35B-A3B, and Qwen3.8 (July 2026; 2.4T parameters, 95B active; 27B distilled Apache variant); QwQ reasoning line. Qwen
  • Z.ai (Zhipu) — GLM-5 (February 2026), GLM-5.2 (June 2026), GLM-5.3 (August 2026, 744B, 40B active), and GLM-5.3-Flash; earlier GLM-4.5/4.6; MIT/Apache licensed. Z.ai
  • Moonshot — Kimi K1.5, K2, K2.5 (multimodal), K2.6, K2.7 Code, and Kimi K3 (July 2026, 2.8T parameters, 104B active)—the largest open-weight model to date. Moonshot
  • ByteDance — Doubao Large Model (Volcano Engine), Seed and Seed-oss, Seedream, Seedance 2.0. Doubao
  • Baidu — ERNIE 4.5 and ERNIE X1 (reasoning). Baidu

Other Regions

  • Korea — LG’s K-EXAONE 2.0 (750B, 37B active, July 2026) sovereign open model; Samsung’s Gauss on-device models.
  • Japan, India, UAE — national AI programs and sovereign-model efforts.
  • Reasoning models (o3, DeepSeek-R1, QwQ, ERNIE X1) spend tokens on deliberation before answering.
  • Open-weight models (Llama, Qwen, DeepSeek, GLM, Kimi, Mistral) enable local deployment and rival closed frontier models.
  • Multimodal and long-context capabilities (one-million-token contexts in DeepSeek V4, Gemini, and Kimi) are now standard.
  • World models (Genie 3, Cosmos 3, Waymo World Model, Happy Oyster) extend the frontier beyond text; see World Models.

Why It Matters

Frontier capability is no longer US-only. China’s DeepSeek, Qwen, GLM, and Kimi families now release open weights that rival US frontier models, and Europe’s Mistral anchors the EU’s sovereignty push.

Sources

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

Related articles

History of Artificial Intelligence Recent Topics in Artificial Intelligence