Kimi K3
Multimodal Kimi model with 1M context and toggleable max-effort thinking for long-horizon agent work
- Released
- Jul 16 2026
- Modality
- Text + Image + Video
- Context
- 1M tokens
- Price in / cached input / out
- $2.85 / $0.28 / $14.25per 1M tokens
Quickstart
OpenAI-compatible: change the base URL, and the setup is complete.
from openai import OpenAI
import os
client = OpenAI(
base_url="https://api.you.com/v1",
api_key=os.environ["YOU_API_KEY"],
)
response = client.chat.completions.create(
model="kimi-k3",
messages=[{"role": "user", "content": "Summarize this support ticket in two sentences."}],
)
print(response.choices[0].message.content)Specifications
Technical details, capabilities, pricing, and data handling.
Model info
- Input Modality
- Text, Image, Video
- Output Modality
- Text
- Context Window
- 1M
- Parameters
- 2.8T total / 104B active
- License
- Kimi K3 License
Capabilities
- Reasoning
- Supported
- Function calling
- Supported
- Structured outputs
- Not supported
- Web search grounding
- Available with You.com Search
Pricing
- Input
- $2.85
- Cached input
- $0.28
- Output
- $14.25
Data handling
- Zero data retention
- Supported
- No training on customer data
- Supported
- Regions
- US
Providers
Access models through multiple providers with automatic routing and built-in failover.
| Provider | ZDR | No training | Region |
|---|---|---|---|
| You Inference | Yes | Yes | US |
| Parasail | Yes | Yes | US |
| Wafer | Yes | Yes | US |
When to use Kimi K3
Guidance based on the specifications and benchmarks on this page.
Best for
- Long-horizon autonomous codingSustains long engineering sessions with minimal oversight, navigating large repositories and orchestrating terminal tools.
- Agentic research and knowledge workBuilt for end-to-end knowledge work such as deep research, dashboards, and multi-step tool use.
- Vision-in-the-loop developmentNative image understanding lets it iterate between code and screenshots for frontend, game development, and CAD work.
Consider alternatives for
- You need peak frontier qualityMoonshot says K3's overall performance still trails the most powerful proprietary models.
- You need low-cost, high-volume inferenceSmaller models such as DeepSeek V4.1 Flash or GPT-OSS 120B fit simple, high-volume workloads at lower cost.
- You need audio input or outputKimi K3 handles text, images, and video; audio is not a supported modality.
Benchmarks
Kimi K3 results alongside comparable models
| Benchmark | Kimi K3Moonshot AI | Qwen3.8 MaxAlibaba | DeepSeek V4 ProDeepSeek | DeepSeek V4.1 FlashDeepSeek | GLM 5.3Z.ai | GPT-OSS 120BOpenAI |
|---|---|---|---|---|---|---|
| GPQA Diamond | 93.5% | 92.6% | 90.1% | 90.9% | Not evaluated | 80.1% |
| Humanity's Last Exam | 43.5% | 43.6% | 37.7% | 36.8% | Not evaluated | 14.9% |
| Humanity's Last Exam (with tools) | 56% | 56.2% | 48.2% | 63.9% | 62.5% | 19% |
Scores are self-reported by each model’s developer, so settings may vary between models. Sources: Kimi K3 (Hugging Face), Qwen3.8 Max (Hugging Face), Qwen3.8 Max (release post), DeepSeek V4 Pro (Hugging Face), DeepSeek V4.1 Flash (Hugging Face), GLM 5.3 (Hugging Face), GPT-OSS 120B (model card).
Give Kimi K3 access to the live web
Ground responses with fresh web information and cited sources using You.com Search, without changing models or adding another integration.
What error do I get installing psycopg2 on Python 3.13?
LLM + You.com Web Search
psycopg2 on Python 3.13
Short answer:
psycopg2 installs fine today. 2.9.10 added 3.13 support in Oct 2024. 12
The error you hit is almost certainly missing build step:
pg_config executable not found
Fix:
pip install -U psycopg2-binary
LLM (LLM only)
When trying to install psycopg2 on Python 3.13, you will most likely hit one of these errors:
1. No pre-built wheel available
pip cannot find a binary wheel because the psycopg2 project hasn’t yet published wheels for Python 3.13.
What you should do
Downgrade Python — use Python 3.12 or earlier until psycopg2 officially supports 3.13.
Explore popular models
Explore all models2.4-trillion-parameter MoE flagship for coding, professional work, multimodal understanding, and long-horizon agentic workflows
Flagship GLM model for long-horizon coding, agents, and complex project delivery
DeepSeek V4.1 Flash model for reasoning and agentic coding
Open MoE flagship with million-token context for coding and long agent runs
Build with Inference Router
Tell us what you run today, and a solutions engineer will show you what it costs on open-weight models.