Updated

A large brushed-silver crescent sculpture rising out of dark navy mist on a dark background, abstract hero artwork for Kimi K3

Moonshot AI

Kimi K3

Moonshot AI's open 2.8T-parameter model for long coding sessions, research and video input.

Chat with Kimi K3 in Super AgentView pricing

What is Kimi K3?

Kimi K3 is Moonshot AI's most capable model, launched in July 2026 with 2.8 trillion parameters (104B active), which Moonshot calls the first open 3T-class model.

It is built for long coding sessions with little oversight, end-to-end research that ends in charts, widgets and dashboards, and native understanding of text, images and video within a 1M-token context window. Moonshot says it still trails Claude Fable 5 and GPT-5.6 Sol overall but beats every other model in its evaluation suite. Moonshot designed K3 to always think, with low, high and max effort.

Provider
Moonshot AI
Context window
1M tokens
Input
Text, Image, Video
Reasoning
Off, low, medium, high, max
Max output
1M tokens
Released
July 2026

How Kimi K3 performs

Published by Moonshot AI

Results Moonshot AI reported for Kimi K3. Each lab runs its own setup, so compare these with other models only inside the same table.

Benchmark results Moonshot AI published for Kimi K3
BenchmarkScoreSetting
GPQA Diamond93.5%max effort
Humanity's Last Exam (full)56.0%with tools, max effort
DeepSWE v1.167.5%Kimi Code harness
Terminal-Bench 2.188.3%Kimi Code harness
BrowseComp91.2%context compaction at 300K
Toolathlon-Verified76.5%max effort
MCP-Atlas84.2%max effort
OSWorld-Verified84.8%max effort
Video-MME90.0%with subtitles, max effort
MMMU-Pro81.6%no tools, max effort

Source: huggingface.co

What you can do with Kimi K3

  • Open weights

    2.8T total parameters, 104B active

  • Native text, image and video understanding

  • Long coding sessions across large repositories

  • Research output with interactive charts and dashboards

  • 1M-token context window

How to prompt Kimi K3

  • Spell out which decisions are yours to make; Moonshot notes K3 can be over-proactive.
  • Attach images or a short video when you want it to reference them.
  • Start a fresh chat with K3 rather than switching to it mid-conversation, since it relies on its earlier reasoning.

Where Kimi K3 fits best

  • Tool-heavy agent runs

  • Large, multi-part projects

  • Reviewing images and video clips

  • Research that ends in a dashboard

How Kimi K3 compares

Kimi K3 compared with other chat models on Flashloop
ModelLMArena textLMArena codingContext windowReasoning
Kimi K31487#181539#111M tokensOff, low, medium, high, max
Gemini 3.1 Pro Preview1487#191522#321M tokensLow, medium, high
GPT 5.6 Sol1485#211534#171.05M tokensOff to max
Gemini 3.7 Flash1490#161520#361M tokensLow, medium, high
Claude Opus 51491#131528#231M tokensOff to max

Good to know before you start

  • Moonshot documents only low, high and max effort, and says K3 can't stop thinking
  • Can make decisions you didn't ask for on ambiguous tasks, Moonshot warns
  • Output can degrade when earlier reasoning is dropped from the conversation

Start chatting with Kimi K3

Chat with Kimi K3 in Super Agent

  1. Select Kimi K3

    Open Super Agent on Flashloop and choose Kimi K3 in the model picker.

  2. Enter your prompt

    Ask for what you need: an answer, a plan, or the images and videos you want the agent to make.

  3. Send and keep going

    Send the message. Kimi K3 answers in Super Agent and runs any generations you asked for.

Questions about Kimi K3

Is Kimi K3 open-weight?

Yes. Moonshot AI publishes the weights under the Kimi K3 License. The model has 2.8 trillion parameters, with 104 billion active per token.

What inputs does Kimi K3 accept?

Text, images and video, with a 1M-token context window.

What reasoning levels does Kimi K3 support?

Moonshot documents low, high and max effort and says K3 always thinks. Super Agent's picker lists off, low, medium, high and max.

How is Kimi K3 different from Kimi K2.6?

Kimi K3 is Moonshot's newer flagship, with a 1M-token window, video understanding built into the model and graded effort levels. Kimi K2.6 is an earlier coding and agent model with a 256K window, a thinking on or off switch, and video input only as an experimental feature of Moonshot's own API.

More chat models on Flashloop

Browse all models