Moonshot AI Frontier Open weight Available in FlowHunt

Kimi K3

Kimi K3 is a top-tier frontier AI model developed by Moonshot AI and released in July 2026. It is an open-weight model: the full weights are published under the Kimi K3 License, and it is also available through Moonshot’s API, which is compatible with OpenAI and Anthropic client formats. The 1M-token context window is large enough to hold entire codebases, lengthy technical documents, or extended agentic sessions without truncation.

Read more
Moonshot AI calls Kimi K3 its most capable model to date and the first open 3T-class model. It is a mixture-of-experts model with 2.8T total parameters and 104B active, built on Kimi Delta Attention and Attention Residuals, with native vision. It is designed for long-horizon coding, agentic knowledge work and reasoning: Moonshot describes it sustaining long engineering sessions, navigating massive repositories and orchestrating terminal tools, and producing research with interactive visualizations and dashboards. Thinking is always on, with the effort configurable between low, high and max.

Benchmarks

Each score appears once, with its rank among tracked models and who measured it.

Compared with Frontier peers on the same scale
  • Kimi K3this model
  • GPT-6.1 Solthis model leads on 0 of 4
  • Claude Opus 5.5this model leads on 0 of 5
Verified independent evaluator, counted in the ranking
Overall

AA Intelligence Index

Independent composite of agent, coding, knowledge and science tests by Artificial Analysis (v4).

Source ↗

Kimi K3 43.6 GPT-6.1 Sol 51.8 Claude Opus 5.5 57.6
#13of 45 Verified
Coding

DeepSWE

113 long-horizon software tasks written from scratch, graded by hidden tests.

Source ↗Also: Self-reported 67.5% (Kimi Code harness)

Kimi K3 68.5% GPT-6.1 Sol No score Claude Opus 5.5 74.2%
#7of 17 Verified

SWE-bench Verified

500 verified GitHub issues. Saturated at the frontier.

Source ↗

Kimi K3 93.4% GPT-6.1 Sol No score Claude Opus 5.5 No score
#5of 22 Verified
Agents

Terminal-Bench 4.0

Multi-step professional tasks done by an agent in a command-line terminal.

Source ↗

Kimi K3 12.6% GPT-6.1 Sol 56.1% Claude Opus 5.5 59.6%
#22of 36 Verified

Humanity's Last Exam

Expert-written questions across 100+ subjects, no tools.

Source ↗Also: Self-reported 43.5% (HLE-Full, no tools)

Kimi K3 46.9% GPT-6.1 Sol 52.9% Claude Opus 5.5 61.4%
#13of 44 Verified

GPQA Diamond

Graduate-level science questions that search cannot answer.

Source ↗Also: Verified 92.9% (Vals AI own implementation)

Kimi K3 93.5% GPT-6.1 Sol No score Claude Opus 5.5 No score
#6of 37 Verified

ARC-AGI-2

Novel visual puzzles that cannot be memorised.

Source ↗

Kimi K3 60.4% GPT-6.1 Sol 94.2% Claude Opus 5.5 93.3%
#19of 27 Verified

What to Use Kimi K3 For

Long-Horizon Coding

Moonshot describes it working with minimal human oversight across long engineering sessions, large repositories and terminal tools, from GPU kernel optimization to compiler development.

Self-Hosted & Private Deployments

Kimi K3 is open-weight, so you can run it on your own infrastructure to keep data in your environment or fine-tune it on proprietary data. It is a very large model, so plan for substantial hardware.

Agentic Knowledge Work

It produces deep research with interactive visualizations, widgets and dashboards, and Moonshot also points to motion design and video editing powered by its native multimodal design.

Multimodal Understanding

It reads text, images and video within the same model, which suits tasks that mix documents, screenshots and code.

Strengths and limitations

Strengths

  • Open weights for self-hosting and fine-tuning
  • Long-horizon coding and agentic work
  • Native multimodality across text, images and video
  • Very long documents and large codebases (1M context)
  • OpenAI- and Anthropic-compatible API

Limitations

  • Kimi K3 License is a modified licence, not an OSI-approved open-source one
  • Very large (2.8T total parameters), so self-hosting needs substantial infrastructure
  • Thinking is always enabled, and multi-turn use requires passing the reasoning content back with each request

Use it in FlowHunt

Use Kimi K3 in any flow, next to your own tools and knowledge sources.

  1. 1Open a flow, or create one from a template
  2. 2Add an AI Agent or Generator component
  3. 3Pick Kimi K3 in the model selector

Related AI Models

About Moonshot AI Beijing, China · Founded 2023 · moonshot.ai

Moonshot AI is a Beijing-based AI company founded in 2023. It develops the Kimi family of models and the Kimi assistant.

moonshot.ai ↗

Browse the Full AI Model Leaderboard

Frequently asked questions