Back to kimi
Moonshot AI /

KI
Kimi K3

open-sourceChat & ReasoningtextimagevideoUpdated July 25, 2026

Kimi K3: Frontier Intelligence & Open Weights Reasoning

Kimi K3 is Moonshot AI's flagship 2.8T parameter open-source mixture-of-experts model, architected specifically for complex abstract reasoning, long-horizon coding, and multimodal intelligence. Built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), Kimi K3 sets a new benchmark for open-weights performance while operating with unprecedented context caching efficiency.


🔬 Architecture & Methodology

Kimi K3 replaces traditional dense attention mechanisms with a hybrid attention-residual architecture engineered for 1M+ token context scaling.

bash
       ┌────────────────────────────────────────────────────────┐
       │                 Output Layer (AttnRes)                 │
       └───────────────────────────▲────────────────────────────┘
                                   │
      ┌────────────────────────────┴───────────────────────────┐
      │  3× MoE Blocks (Stable LatentMoE + Gated MLA + KDA)    │
      └────────────────────────────▲───────────────────────────┘
                                   │
      ┌────────────────────────────┴───────────────────────────┐
      │        Kimi Delta Attention (KDA) Base Module          │
      └────────────────────────────▲───────────────────────────┘
                                   │
      ┌────────────────────────────┴───────────────────────────┐
      │       Shared Experts (2) + Routed Experts (16/896)     │
      └────────────────────────────────────────────────────────┘

Core Innovations

  1. Kimi Delta Attention (KDA): A state-space hybrid operator that reduces KV-cache memory overhead by 72% while preserving long-range attention accuracy up to 1M tokens.
  2. Stable LatentMoE: Features 896 total fine-grained experts, dynamically routing 16 active experts per token for ultra-high memory bandwidth efficiency.
  3. Attention Residuals (AttnRes): Replaces standard additive residual connections with a learned linear gate ($\alpha$) across residual blocks to prevent signal degradation in deep networks.

📊 Benchmark Performance

Kimi K3 demonstrates frontier-grade performance across complex reasoning, competitive coding, and math evaluation suites:

Specification Table
BenchmarkTask DomainKimi K3 ScorePrior SOTA OpenStatus
ARC-AGI-2Abstract Reasoning77.1%31.1%Verified
SWE-bench VerifiedSoftware Engineering68.4%52.8%Verified
MATH-500Competition Mathematics94.2%90.1%Verified
HumanEvalCode Generation92.6%89.4%Verified
MMMUMultimodal Reasoning78.9%73.5%Self-Reported

💡 Note: Context caching is automatically enabled for requests matching repeating system prompts or long documents, reducing latency by up to 10× on repeat calls.


🚀 Get Started

The examples below require Python 3.9+ and the OpenAI Python SDK. Install the SDK and initialize the client once to start issuing reasoning completions.

1. Installation

bash
python3 -m pip install --upgrade 'openai>=1.0.0'

2. Environment Setup

bash
export MOONSHOT_API_KEY="your-moonshot-api-key"

3. Basic Completion Call

python
import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MOONSHOT_API_KEY"],
    base_url="https://api.moonshot.ai/v1",
)

completion = client.chat.completions.create(
    model="kimi-k3",
    messages=[
        {
            "role": "system",
            "content": "You are Kimi K3, a frontier AI assistant specialized in complex reasoning and software architecture."
        },
        {
            "role": "user",
            "content": "Explain Kimi Delta Attention (KDA) in one concise sentence."
        }
    ],
    temperature=0.3,
)

print(completion.choices[0].message.content)

4. cURL API Example

bash
curl https://api.moonshot.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $MOONSHOT_API_KEY" \
  -d '{
    "model": "kimi-k3",
    "messages": [
      {
        "role": "user": "Write a Python script to compute prime numbers up to N."
      }
    ]
  }'

🔗 Official Resources

Key Features

Built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes)

Feature 01

Stable LatentMoE framework efficiently activates 16 out of 896 experts

Feature 02

1M-token context window with automatic context caching

Feature 03

Native visual understanding supporting images and video files

Feature 04

Always-on thinking mode with configurable reasoning effort

Feature 05

You might also want to compare

Verified Sources

Tags

moethinking-modereasoningcodingvision

Model Specs

open-source

Parameters

2.8T

Context Window

1M

Tier

K3

License

Other/Custom

Deployment

api-onlyself-hostable

Resources & Links

Lineage

Model Family

Part of the kimi family

Only release in this line currently tracked.

Predecessor

Kimi K2.6

Compare Specs

Compare parameters, context windows, modalities, and benchmark scores of this model side-by-side with others.

Compare Model