Kimi
An AI assistant for everyone: search, reading, Q&A, writing, and agent tasks.
2026.07.17 · OFFICIAL LAUNCH BRIEF
From answering, to finishing.
Kimi is Moonshot AI's family of AI assistants and agent products. The new Kimi K3 brings long-horizon coding, knowledge work, visual understanding, and deep reasoning into one open-weight flagship blueprint.
LIVE NOW · FULL MODEL WEIGHTS PLANNED BY 2026.07.27
OPEN FRONTIER INTELLIGENCE03 / K3
2.8Ttotal parameters
1Mtoken context
16 / 896active / total MoE experts
Nativenative vision
01
KIMI, BEYOND CHAT
Kimi is more than a model name.
Kimi is the product layer people use; K3 is the model layer powering a new generation of task execution. Together they span conversation, research, code, and enterprise collaboration.
Kimi Work
Built for knowledge work: deep research, documents, slides, spreadsheets, and interactive dashboards.
Kimi Code
A coding agent for terminals and IDEs that reads and writes code, invokes tools, and executes long-running tasks.
Kimi API
For developers and organizations, use kimi-k3 to bring K3 into your products.
02
WHAT K3 CHANGES
A model designed for long-horizon tasks.
K3 is not just about answering one more question correctly. It is designed to understand its environment, call tools, inspect results, and finish complex work with less human supervision.
01 / AGENTIC CODING
From generating code once to working through an engineering project.
It can navigate large repositories, run terminal tools, and iterate on frontend, game, and CAD work using visual feedback.
$ kimi --model k3
● reading 1,284 files
● running tests
task complete
02 / NATIVE VISION
Vision enters the task loop
Text, images, and video are understood by one model, so it can see its results and refine them in the loop.
03 / KNOWLEDGE WORK
From a pile of sources to a usable deliverable
Built for research, consulting, documents, spreadsheets, presentations, and interactive data stories.
04 / DIGITAL CREATION
Code, 3D, and moving-image creation converge.
Official demos include a browser-based 3D open world, motion explainers, and multi-clip video editing.
03
UNDER THE HOOD
Behind 2.8T, there is more than scale.
K3 introduces new attention and depth-wise information-flow designs and scales MoE sparsity to 896 experts. Moonshot reports roughly 2.5× better overall scaling efficiency than K2.
1M
CONTEXT
CONTEXT
KDAlong-sequence attention
AttnRescross-depth retrieval
LatentMoE16 / 896 experts
K3
OUTPUT
OUTPUT
KDA
Kimi Delta Attention provides a more efficient attention foundation for very long sequences.
AttnRes
Instead of uniformly accumulating every layer, it selectively retrieves representations from different depths.
MXFP4 / 8
Quantization-aware training starts at SFT, using MXFP4 weights and MXFP8 activations.
04
OFFICIAL EVALUATION
Read the scores—and the test conditions.
These K3 max results come from Kimi's launch blog. Benchmarks use different agent harnesses, tools, and sampling settings, so treat them as a capability profile—not an absolute ranking.
KIMI K3 / MAX REASONINGChoose a capability
Temperature 1.0 and top-p 1.0; other conditions vary by benchmark. Internal benchmarks are not shown here.
05
AVAILABLE NOW
What can you do with K3 today?
Use it directly in Kimi and Kimi Work, select K3 in Kimi Code, or call kimi-k3.
Python / OpenAI SDK
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["MOONSHOT_API_KEY"],
base_url="https://api.moonshot.cn/v1",
)
response = client.chat.completions.create(
model="kimi-k3",
messages=[{"role": "user", "content": "Complete this research task"}],
reasoning_effort="max",
)KIMI API LAUNCH PRICING
Price per million tokens
Cache-hit input$0.30
Cache-miss input$3.00
Output$15.00
Pricing is from the K3 launch blog. Check the API platform for current rates.
Open Kimi API PlatformREAD BEFORE USE
Kimi also documents K3's limitations.
Do not switch models mid-session
K3 is sensitive to complete thinking history; switching from another model mid-session can make quality unstable.
It can be overly proactive
When intent is ambiguous, K3 may make unexpected decisions on the user's behalf. Set explicit behavioral boundaries.
It does not surpass every closed frontier model
Moonshot acknowledges a noticeable user-experience gap between K3 and the strongest proprietary models.
PRIMARY SOURCES
Sources
Content verified on July 17, 2026. Primary Kimi and Moonshot AI sources are used wherever possible.
THE NEXT TASK IS YOURS