Kimi K3 is live on Featherless

Moonshot just dropped Kimi K3, and it is the first open-weight model to size up at nearly 3 trillion parameters. See why Washington DC is scared of this model, live on Featherless.

What changed from K2

K2 was about 1 trillion total parameters and activated 32 billion per token. K3 scales that to 2.8 trillion total and roughly 104 billion active, selecting 16 of 896 experts for each token instead of running the entire model. Moonshot claims their new innovative architecture (such as their Kimi Delta Attention mechanism) offers 2.5× better scaling efficiency than K2.

Kimi K3 also continues Moonshot’s push into native multimodality. Like K2.5, it can take images directly as input and reason about what it sees, with K3 upgrading the vision system with Moonshot’s new vision encoder, MoonViT-V2.

The numbers

In short, the numbers are great. Kimi K3 runs toe-to-toe with the best closed models on Terminal-Bench 2.1, scoring 88.3 against Claude Fable 5’s 88.0 and GPT-5.6 Sol’s 88.8 – and these are against closed weight models. Kimi K3 also fights hard on GPQA Diamond (a standard benchmark for difficult scientific reasoning) scoring 93.5 compared with Claude Fable 5 at 92.6 and GPT-5.6 Sol at 94.1. You can read Moonshot’s benchmarks here. If you’re interested in benchmarks from 3rd parties, check out the leaderboard from Arena AI. This is the part Washington is paying attention to: performance this close to the frontier is now coming from an open-weight model out of China. Regardless, our mission at Featherless is simple: make the best open models available to everyone.

How to run Kimi K3

Currently on Featherless, Kimi K3 serves at: 32K context, FP8 quantization, vision enabled, but we’re working hard to expand these numbers so you can use these models with larger context.

Kimi K3 runs on the same LLM calling API developers all know and use, the OpenAI-compatible API. It’s the same as our other thousands of models we host on our site.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.featherless.ai/v1",
    api_key="YOUR_FEATHERLESS_KEY",
)

response = client.chat.completions.create(
    model="moonshotai/Kimi-K3",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": "What do you see in this image?"
                },
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "https://example.com/image.jpg"
                    }
                }
            ]
        }
    ],
)

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

It’s that simple.

And, as is the same with all of your api calls on featherless, no logs, so your prompts and completions stay yours.

So what are you waiting for?

Try it out today – http://featherless.ai/models/moonshotai/Kimi-K3

Going beyond

Are you interested in running Kimi K3 at even larger scales? Are your unpredictable per-token bills exploding? Check out our dedicated gpu clusters, where our team can help tune the full inference stack exactly for your needs, including quantization, batching, fine-tuning, and whatever else you need. We have GPUs, and you need compute. Learn more below, and feel free to schedule a call, where we’ll help tackle your biggest AI problems.

Get Dedicated GPU at $7.6k/month