Moonshot AI

Moonshot AI

Overview

Moonshot AI is a Beijing-based Chinese AI company founded in 2023 by Yang Zhilin, focused on long-context, agentic large language models. The company's Kimi product line — initially distinguished by industry-leading context length when it launched in 2023 (~200K tokens at a time when competitors offered 32K) — has evolved into one of the most capable open-weight agentic model families in 2026. The current flagship, Kimi K2.6, is a native multimodal agentic model built on a 1-trillion parameter Mixture-of-Experts architecture with 32B parameters activated per token, released April 20, 2026 under a Modified MIT license.

Moonshot's distinctive technical pitch in 2026 is "Agent Swarm" — a multi-agent orchestration architecture built into K2.6 that scales to 300 domain-specialized sub-agents executing up to 4,000 coordinated steps in a single autonomous run. Combined with K2.6's 262K context window, this positions Moonshot as one of the leading non-Western frontier agentic providers, with benchmark results that compete directly with the latest GPT, Claude, and Gemini releases on coding (SWE-Bench Pro, Humanity's Last Exam) at substantially lower price.

Key Details

  • Founded: 2023
  • CEO / Founder: Yang Zhilin
  • Headquarters: Beijing, China
  • Distribution Model: Open-weight via Hugging Face / DeepInfra; Kimi consumer chat product; API platform
  • Strategic Positioning: Long-context, open-weight, agentic LLMs
  • Website: https://moonshot.ai / https://kimi.com

Current Models

  • [[Kimi K2.6]] — Current flagship; 1T-parameter MoE with 32B activated per token; native multimodal; 262K context; Agent Swarm architecture (300 sub-agents, 4,000 coordinated steps); Modified MIT open-weight license; released April 20, 2026
  • Kimi K2 / K2.5 — Earlier-generation models in the same lineage
  • Kimi (consumer chat) — Free Chinese-market chat product built on the K-series models

Recent Developments

  • Kimi K2.6 Launch (April 20, 2026): 1T-parameter MoE released as open-weight under Modified MIT. Agent Swarm architecture scales to 300 domain-specialized sub-agents executing up to 4,000 coordinated steps in a single run — up from 100 sub-agents and 1,500 steps in K2.5.
  • Benchmark Performance:
    • 58.6 on SWE-Bench Pro (vs. 57.7 for GPT-5.4)
    • 54.0 on Humanity's Last Exam (HLE-Full) with tools — leading every model in the comparison, including GPT-5.4 (52.1), Claude Opus 4.6 (53.0), and Gemini 3.1 Pro (51.4)
  • 262K Context: Extension of Moonshot's long-context heritage; supports the agentic workflows and long-horizon coding that the Agent Swarm architecture enables.
  • Open-Weight Strategy: Continues Moonshot's commitment to releasing flagship models under permissive open licenses — a deliberate competitive position against Western frontier labs (closed) and a complement to DeepSeek, Qwen, and the broader Chinese open-weight push.

Why They Matter

Moonshot is the leading Chinese frontier lab on agentic and long-context capability — and its Agent Swarm architecture is one of the most ambitious public attempts to scale multi-agent orchestration as a primary capability rather than an external orchestration layer. K2.6's benchmark results — leading the field on Humanity's Last Exam and matching GPT-5.4 on SWE-Bench Pro at open-weight pricing — make a credible case that frontier capability is becoming commoditized through Chinese open-weight releases. Combined with DeepSeek V4-Pro (1.6T MoE, MIT) and Alibaba Qwen 3.5 (397B-A17B), the Chinese open-weight stack in early-to-mid 2026 represents the most serious sustained challenge to Western closed-source frontier labs since the rise of GPT-4. For enterprises evaluating open-weight agentic deployment, Moonshot is the lab pushing the multi-agent orchestration frontier hardest.

Last Updated

May 7, 2026

→ Back to Companies