With the release of the full Kimi K3 model weights, we have completed its GPU-based deployment and integrated it into our existing API service. Developers can now access Kimi K3 directly through an OpenAI-compatible API—without downloading massive model weights, setting up an inference cluster, or maintaining service processes themselves.
This release offers two service tiers:
kimi-k3: Standard versionkimi-k3-fast: Fast, high-speed version
Both tiers use the same Kimi K3 weights and offer the same context capabilities. The only differences are response speed and price: the Fast version responds more quickly and costs 1.5 times as much as the standard version.
kimi-k3-fast is a high-speed service tier offered by our platform, not a separate set of model weights released by Moonshot AI.
What Is Kimi K3?
Kimi K3 is an open-weight, natively multimodal agent model released by Moonshot AI. According to its official model card, it uses a mixture-of-experts architecture with 2.8 trillion total parameters, activates approximately 104 billion parameters per token, and supports a context window of around 1 million tokens.
The model consists of 896 routed experts, with 16 experts selected for each token. It also incorporates architectural innovations such as Kimi Delta Attention, Attention Residuals, and Stable LatentMoE, and is designed primarily for long-horizon coding, complex reasoning, knowledge work, and agentic tasks.
It is important to distinguish between “open weights” and “unconditionally open source.” Kimi K3’s code and model weights are released under the separate Kimi K3 License, rather than the Apache 2.0 or MIT License. Teams planning large-scale commercial deployments should still review and comply with the official licensing terms.
What’s the Difference Between the Standard and Fast Versions?
| Comparison |
kimi-k3 |
kimi-k3-fast |
| Service Tier |
Standard |
Fast |
| Model Weights |
Kimi K3 |
Same as the standard version |
| Context Capabilities |
Same |
Same |
| Inference Speed |
Standard speed |
Faster |
| Price |
Standard price |
1.5× the standard version |
| Recommended Use Cases |
Everyday API calls and batch workloads |
Real-time interactions and latency-sensitive workloads |
The Fast version does not switch to a smaller model or reduce the context window to achieve higher speeds. It is best understood as a high-speed service tier for the same model.
Note that “1.5 times the price” does not mean a fixed 1.5× increase in speed. Actual response times also depend on input and output length, context size, concurrency, and task complexity.
Which Version Should You Choose?
If your application does not have particularly strict response-time requirements, we recommend starting with kimi-k3. It is better suited for the following scenarios:
- Blog posts, copywriting, and long-form content generation
- Code analysis, code review, and development testing
- Document organization, summarization, and knowledge-based Q&A
- Background batch-processing tasks
- High-volume applications where cost control is important
If users need to wait for results in the foreground, or if the product experience is particularly sensitive to response times, consider choosing kimi-k3-fast:
- Online chat and AI-powered customer support
- Interactive assistants in IDEs, workbenches, and similar environments
- Agents and multi-step tool calling
- Automated workflows that require rapid feedback
- User-selected, high-priority tasks
A practical approach is to use the standard version by default, then switch to the Fast version when the user selects “Fast Generation” or when a task enters a high-priority workflow. This helps keep overall costs under control while providing a better experience for latency-sensitive users.
Using the OpenAI-Compatible API
If your project already uses the OpenAI SDK, you typically only need to replace the API base URL, API key, and model name.
cURL Example
curl "https://caicaiapi.cloud/v1/chat/completions" \
-H "Authorization: Bearer your api key" \
-H "Content-Type: application/json" \
-d '{
"model": "kimi-k3",
"messages": [
{
"role": "user",
"content": "Please help me analyze the architecture of this project and provide the three highest-priority optimization suggestions."
}
],
"stream": true
}'
To use the Fast version, simply change the model name:
Python Example
from openai import OpenAI
client = OpenAI(
api_key="Your api key",
base_url="https://caicaiapi.cloud/v1",
)
response = client.chat.completions.create(
model="kimi-k3",
messages=[
{
"role": "user",
"content": "Please help me analyze the architecture of this project and provide the three highest-priority optimization suggestions.",
}
],
)
print(response.choices[0].message.content)
Switching to the Fast version likewise requires changing just one line:
Both tiers use the same API format, so your application does not need to maintain two separate integration paths. You can dynamically select the model based on configuration, user tier, or task priority.
What Should You Keep in Mind When Using Kimi K3?
1. Preserve the Complete Reasoning History
Kimi K3 returns a reasoning_content field. In multi-turn conversations and tool-calling workflows, the official recommendation is to include the complete Assistant message from the previous turn in messages exactly as returned—including reasoning_content and tool_calls—rather than retaining only the final content.
If the reasoning history is lost along the way, the stability of long-running and agentic tasks may be affected.
2. Avoid Switching Models Mid-Conversation
For long conversations initiated with another model, switching directly to Kimi K3 midway is not recommended. A more reliable approach is to start a new conversation or first summarize the existing context clearly before handing it over to Kimi K3.
3. Refer to the Current API Documentation for Multimodal Capabilities
Kimi K3 has native multimodal capabilities, but the supported input types may vary depending on the inference engine and service configuration. Please refer to our Model Marketplace and API documentation for the capabilities currently available.
Frequently Asked Questions
Is the Fast Version More Capable?
No. Both tiers use the same model weights and offer the same context capabilities. The Fast version’s primary advantage is faster inference.
Is the Fast Version Always 1.5× Faster?
No. The Fast version is priced at 1.5 times the standard version, but the speed improvement is not a fixed multiplier. Actual performance depends on the request content and current service load.
Can I Use Both Versions in the Same Project?
Yes. You can use kimi-k3 for regular tasks and kimi-k3-fast for real-time interactions or high-priority tasks. Simply switch the model parameter in the request.
How Many Credits Does It Cost?
Please refer to the Model Marketplace for the current credit pricing of both tiers. The Fast version is always priced at 1.5 times the standard version.
Final Thoughts
The GPU-deployed Kimi K3 model is now available in both Standard and Fast tiers.
Choose kimi-k3 when you want to control costs or handle routine and batch workloads. Choose kimi-k3-fast when response speed and a real-time interactive experience matter more.
Both tiers offer the same capabilities and API format, allowing you to switch between them seamlessly within the same project. We will continue improving inference stability, the API experience, and the accompanying documentation.
References
