Kimi K3 — Biggest, most efficient model from Moonshot AI
Kimi K3 is Moonshot AI's flagship reasoning model built for long-horizon coding, agentic knowledge work, and native multimodal understanding at frontier scale.
Frontier intelligence, built to scale
A 2.8-trillion-parameter architecture designed for the hardest, longest-running work.
Long-horizon coding
Sustains multi-day engineering sessions across massive codebases with minimal human supervision, coordinating terminal tools end to end.
Native multimodal reasoning
Understands text, image, and video in one architecture, with "vision-in-the-loop" iteration against live screenshots.
Built to scale
Kimi K3 is one of the largest AI models ever built and its Kimi Delta Attention and Attention Residuals design ensures that extra scale converts into smarter, faster answers.
Deep agentic capability
Orchestrates 20+ concurrent subagents across research, tool use, and multi-step production workflows.
Experience The Brilliance
Kimi K3 makes complex and time-consuming tasks simple and as fast as pressing a button.
Built for Work That Used to Need Supervision
Kimi K3 checks its own work as it goes, sustaining long-running engineering tasks with minimal oversight and navigating large repositories on its own. Ask it to review and analyze complex documents to create advanced relations and it condenses weeks of work into hours.
See and Understand Complex Visual Information
K3's native multimodal architecture unifies text, image, and video understanding, so it can iterate directly against screenshots and live visual feedback. That "vision-in-the-loop" workflow is what makes it effective at frontend engineering, game development, and CAD.
Reasoning You Can Dial Up When It Matters
K3 always reasons before answering, and you control how hard it thinks with a single reasoning_effort setting: low, high, or max. Save max effort for the hardest problems, and drop to low or high when speed and cost matter more.
Everything Kimi K3 Brings
Across coding, knowledge work, and multimodal reasoning, K3 raises the bar in ways that add up fast.
Long-horizon coding
Sustains multi-day autonomous engineering sessions across huge repos with minimal supervision.
1M-token context
Automatic caching keeps long sessions fast and affordable. No manual cache management required.
Reasoning effort
Dial thinking between low, high, and max depending on task difficulty and cost.
Native vision-in-the-loop
Iterates directly against screenshots for frontend, game dev, and CAD work.
Video understanding
Analyzes and edits video natively, from motion graphics to frame-accurate cuts.
Frontier-scale MoE
2.8 trillion parameters, 16-of-896 active experts, ~2.5x the scaling efficiency of Kimi K2.
Agent orchestration
Coordinates 20+ concurrent subagents across research, terminal tools, and multi-step workflows.
Tool-calling APIs
Strict JSON Schema outputs, dynamic tool loading, and required tool-choice control.
Open-weight
Full weights ship July 27, 2026, freely downloadable for self-hosting.
Have questions?
We have answers!
Kimi K3 is Moonshot AI's flagship model with a 2.8-trillion-parameter Mixture-of-Experts model and a 1-million-token context window with native multimodal (text, image, video) understanding. It's the largest open-weight model released to date.
Kimi K3 is open-weight, not open-source in the strict sense. Moonshot is releasing the trained weights (by July 27, 2026) under a Modified MIT license, but not the training data or training code.
Via the official API: $3.00 per 1M input tokens (cache miss), $0.30 per 1M input tokens (cache hit), and $15.00 per 1M output tokens. But you can use it at platforms like Imagine ComputerImagine Computer for $9/month on a yearly plan or $13/month on a monthly plan.
Independent benchmarking from Artificial Analysis ranks K3 4th out of 189 models on its Intelligence Index. Behind Claude Fable 5 and GPT-5.6 Sol, but ahead of Claude Opus 4.8, GPT-5.5, Claude Sonnet 5, and GLM-5.2. It also debuted at #1 on LMArena's Frontend Code Arena.
Moonshot flags two known behaviors: K3 can be "excessively proactive," making decisions on your behalf when instructions are ambiguous, and it's sensitive to thinking history being dropped mid-session. Both are worth constraining explicitly in your system prompt if precision matters.
Try Kimi K3 in Imagine Computer
We've added Kimi K3 to Imagine Computer's AI Chat so you can access Moonshot AI's flagship 2.8T reasoning model directly from your workspace. No separate API key required.