Skip to main content

Kimi K3.1 leaks: three reasoning modes, million-token context

Kimi K3.1 leak: what's inside Moonshot's next flagship

Moonshot AI is preparing to launch Kimi K3.1, and if a new leak is accurate, it could land as early as next month — October 2026. The tip comes from X user @MaxForAI, who dug through internal JSON and API responses from Moonshot and spotted model names like k3d1-agent, along with configuration switches for Agent mode and other features.

Three reasoning levels, a million tokens, and Swarm

According to the leaked configuration files, K3.1 may support several new capabilities:

  • Three reasoning intensity levels — Low, High, and Max
  • An "Extra Long" context option that stretches up to 1 million tokens
  • Agent mode for autonomous task execution
  • Swarm multi-agent collaboration
  • Task modes such as search and batch processing

If those settings are real, K3.1 would push Kimi further into long-context work, agent teamwork, and task scheduling — three areas where AI labs are racing right now.

The K3 benchmark: 2.8 trillion parameters

For context, Moonshot's current flagship Kimi K3 launched in July 2026. Officially, it packs 2.8 trillion parameters, native visual capabilities, and a 1 million token context window. Under the hood, it uses architectural tricks like Kimi Delta Attention, Attention Residuals, and sparse MoE.

So what exactly will K3.1 upgrade on top of that? Will it still deserve the "flagship" label? Those answers will have to wait until Moonshot makes an official announcement. For now, the leak gives us a rough sketch — and a reason to keep an eye on October.

Key Points

  • Kimi K3.1 is rumored to launch in October 2026, per a leak from @MaxForAI on X.
  • The model reportedly offers three reasoning levels: Low, High, and Max.
  • An "Extra Long" context option may support up to 1 million tokens.
  • Leaked configs also point to Agent mode, Swarm multi-agent collaboration, and search/batch task modes.
  • The previous flagship, Kimi K3, launched in July 2026 with 2.8 trillion parameters and a 1 million token context window.