How to Use Kimi K3: Moonshot AI’s 2.8T Open-Weight Model

Moonshot AI’s Kimi K3 is a 2.8-trillion-parameter open-weight mannequin constructed with a Combination-of-Specialists structure. It prompts solely a small fraction of its parameters per token, serving to cut back inference prices whereas delivering sturdy coding and agentic efficiency.

K3 combines near-frontier capabilities, open weights, and decrease API pricing, making it an attention-grabbing various to proprietary fashions. On this article, we’ll discover how K3 works and present you two easy methods to make use of it: by means of Kimi’s app or the Fireworks AI API.

What Makes K3 Completely different?

Two selections in structure are notable: 

• Kimi Delta Consideration (KDA): A hybrid linear consideration method created to render the context window of 1M tokens possible as an alternative of quadratic. 
• Consideration Residuals: A whole substitute for traditional residual connections that Moonshot claims result in extra constant outcomes because the mannequin grows bigger. 

K3 makes use of a sparse Combination-of-Specialists structure that prompts solely a small fraction of its parameters for every token. This offers the mannequin an enormous parameter rely whereas conserving inference comparatively environment friendly in contrast with different trillion-parameter fashions. Early benchmarks counsel that K3 can compete with superior proprietary fashions on coding and agentic duties, with notably sturdy efficiency on long-context coding and blind coding evaluations.

The way to Entry K3 and What it Prices? 

There are two methods to entry K3, and every is suited to a distinct use case.

• Kimi App: Accessible on iOS and Android by means of Kimi.com, in addition to by means of the Kimi Work desktop software. No developer account is required, and you can begin with the free tier, which comes with utilization limits. It’s possibility for those who merely wish to discover the mannequin and see the way it performs.
• API Entry: Accessible by means of the Kimi net platform and Fireworks AI. This feature makes use of pay-per-token pricing and is healthier fitted to builders who wish to combine K3 into their very own functions or workflows.

Fireworks affords three serving paths, relying on the way you wish to stability latency and price:

Cached enter is priced at roughly one-tenth the price of non-cached enter. At scale, it’s due to this fact value structuring your prompts, together with system prompts and reusable context, to benefit from caching.

For many use circumstances, Commonplace is the very best start line. Select Quick when latency is vital, or Precedence whenever you want larger reliability.

If you happen to choose to self-host K3, Moonshot recommends a supernode with not less than 64 accelerators to create a high-speed setting the place the specialists can function inside a single area. This can be a considerably bigger infrastructure dedication than self-hosting fashions within the 700B–1.6T parameter vary.

Now that we all know the whole lot concerning the mannequin, let’s give it a attempt:

Job 1: Speaking to K3 Instantly (No API Key Wanted) 

The quickest option to expertise K3 is thru Kimi’s client interface. You need to use the Kimi app on iOS and Android, entry it by means of kimi.com, or use the Kimi desktop app for work. No developer account or billing particulars are required. You possibly can merely open the chat interface and begin experimenting.

For a extra significant check, don’t simply ask K3 questions on basic information. As a substitute, give it a small mission or ask it to construct a easy system. That’s the place K3’s coding and agentic capabilities actually come into play.

Enter Immediate:

Construct a single-file HTML web page for a typing pace check — it ought to present a random sentence, begin a timer on the primary keystroke, and show words-per-minute and accuracy once I end typing. Make it look clear, not default-Bootstrap. Then clarify the trickiest a part of the logic in two sentences.

Output:

My Remark:

• Is it using its cognitive considering expertise? K3 is at all times working in a “considering state” which suggests it is possible for you to to see it planning its code earlier than “spitting out” the HTML. 
• 100% truthful is what we have to decide whether or not the voiceover = code. This manner, we are able to make certain that it doesn’t simply repeat some phrase it has already heard earlier than.  
• Simply use the visible enter to Geek. Put a screenshot of the developed web site and ask Geek to investigate its margin/distinction. This may show you how to perceive the principle advantages of the system.   

The complete course of may be executed in lower than 5 minutes, which provides you with an understanding of its capabilities in coding and visible notion. 

Job 2: Calling K3 through the Fireworks API 

To construct something with, be it a script, a pipeline or an agent, programmatic entry is required. On this case, Fireworks AI gives K3 serverless service which is OpenAI-compatible by means of accounts/fireworks/fashions/kimi-k3. 

Step 1: Generate a key. Try this from API keys settings of Fireworks and export it 

export FIREWORKS_API_KEY="your-api-key"

Step 2: Relatively than a boring “howdy world” perform making K3 produce one thing in a single go together with its huge 1M-token context window, present it with a picture and inform it to interpret the picture and create one thing coherent from this interpretation in a single go as effectively. This serves as an appropriate imitation of what actual company is about (sense → assume → create product).

import requests
import json

url = "

payload = {
    "mannequin": "accounts/fireworks/fashions/kimi-k3",
    "max_tokens": 4096,
    "top_k": 40,
    "presence_penalty": 0,
    "frequency_penalty": 0,
    "messages": [
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": (
                        "Look at this image. First describe what's in it in one paragraph. "
                        "Then write a 4-line haiku inspired by it. "
                        "Return your answer as JSON with keys 'description' and 'haiku'."
                    )
                },
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "
                    }
                }
            ]
        }
    ]
}

headers = {
    "Settle for": "software/json",
    "Content material-Sort": "software/json",
    "Authorization": f"Bearer "
}

response = requests.put up(
    url,
    headers=headers,
    information=json.dumps(payload)
)

consequence = response.json()

print(consequence["choices"][0]["message"]["content"])

Output:

My Remark:

• Is the JSON format clear, or does K3 cover it in feedback? (You need to know that earlier than deciding to create a pipeline to course of the output).
• How does the standard of the haiku evaluate to the standard of the factual description – is the “artistic” half nearly as good because the half about “notion”?
• Exchange the max_tokens and picture URL with the screenshot of your personal dashboard or your personal sketch and have it created the specification.

How K3 Performs Towards Different AI Fashions?

K3’s efficiency can solely be assessed within the context of different fashions within the area:

• On the Synthetic Intelligence Evaluation Index, K3 is ranked carefully behind GPT-5.6 Sol Max and Claude Fable 5, being only some factors behind the above-mentioned fashions as an alternative of falling considerably behind.
• Within the front-end coding blind evaluation, K3 got here first forward of Claude Fable 5 within the unbiased developer check. 
• Based on Moonshot’s coding evaluation suite and agentic benchmark, K3 outperformed Claude Opus 4.8 and GPT-5.5, whereas falling behind Fable 5 and GPT-5.6 Sol.
• When it comes to the mannequin measurement, K3 is round 75% larger than DeepSeek’s V4 Professional mannequin, being the broadly used open mannequin earlier than K3’s launch, which means that K3 is the most important open-weight mannequin in historical past and never simply moonshot’s mannequin.

The ethical is to remember the fact that K3 excels primarily in programming and long-horizon company duties and never in the whole lot generally. So, it is best to run your personal eval suite concentrating on your particular activity as an alternative of counting on the leaderboard rating alone. 

Conclusion 

K3 is without doubt one of the first fashions in its class that you would be able to obtain and self-host. From an infrastructure perspective, although, Moonshot recommends having 64+ accelerators accessible to run it successfully.

If you wish to discover K3 with out that infrastructure, the Kimi app and Fireworks API are the simplest entry factors. The app enables you to check the mannequin’s capabilities, whereas the API makes it sensible to combine K3 into real-world functions.

Each choices took lower than ten minutes to arrange, and the outcomes had been constant. K3 demonstrated sturdy reasoning earlier than taking motion and dealt with photographs as naturally as different types of enter.

Continuously Requested Questions