Kimi K3

Kimi K3 is a general-purpose open-weight model served on Sarvam infrastructure. It accepts text, images, and video and has a 1,048,576-token context window.

At a glance

Key capabilities

Quickstart

from sarvamai import SarvamAI
client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY")
response = client.chat.completions_v2(
    model="kimi-k3",
    messages=[
        {
            "role": "user",
            "content": "Plan a three-step approach to migrate a monolith to microservices.",
        }
    ],
    max_tokens=4000,
)
print(response.choices[0].message.content)

Supported parameters

model: "kimi-k3" on Chat Completion V2 and Responses. Full request/response schemas live on those API reference pages; this table covers what’s actually accepted for this model.

Reasoning

Thinking is always on and can’t be turned off. Effort levels: low, high, max (default max). Set them with reasoning_effort on chat or reasoning.effort on Responses.

Read the reasoning trace from choices[].message.reasoning_content on chat. On Responses, output starts with a reasoning item; read its content[0].text (summary is empty). Reasoning tokens are billed as output tokens and count against max_tokens / max_output_tokens.

On Responses, when the model calls tools, output has one reasoning item followed by the function_call items for that turn (sometimes with a message item between them).

Image input

Send images inline as base64 data URIs in an image_url content part on chat, or an input_image part on Responses. Supported formats are PNG, JPEG, WebP, GIF, BMP, and TIFF. Images sent by file_id are not supported.

import base64
with open("chart.png", "rb") as f:
    data_uri = "data:image/png;base64," + base64.b64encode(f.read()).decode()
response = client.chat.completions_v2(
    model="kimi-k3",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Summarise this chart."},
                {"type": "image_url", "image_url": {"url": data_uri}},
            ],
        }
    ],
    max_tokens=2000,
)

Images are resized to about 12.8 megapixels. Images above about 179 megapixels are rejected.

Video input (beta)

Send a video inline as a base64 data URI in a video_url content part on Chat Completion V2. Remote video URLs return 400, and the Responses API doesn’t accept video.

import base64
with open("clip.mp4", "rb") as f:
    video_uri = "data:video/mp4;base64," + base64.b64encode(f.read()).decode()
response = client.chat.completions_v2(
    model="kimi-k3",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Describe what happens in this clip."},
                {"type": "video_url", "video_url": {"url": video_uri}},
            ],
        }
    ],
    max_tokens=2000,
)

Audio input is not supported and returns 400.

Tool calling

Pass OpenAI-compatible tools. tool_choice accepts "auto", "none", "required", or a named function ({"type": "function", "function": {"name": "..."}}). The model can return several tool calls in one turn; set parallel_tool_calls: false to get at most one.

Multiple choices

n > 1 needs temperature greater than 0. Sending n > 1 with temperature: 0 returns 400.

Known limitations

• Video input is beta and reliable only for very short clips (see Video input).
• Reasoning and the final answer share the max_tokens / max_output_tokens budget.
• A stop string can match inside the thinking, which ends the response early with content: null. Use stop strings that won’t appear in reasoning.
• Set max_tokens explicitly. When it’s omitted, generation isn’t capped at a small default and long answers can run to tens of thousands of tokens.

Limits and errors