Skip to main content

Kimi K2.6 / K2.7 Code Understanding

Kimi is Moonshot AI's family of natively multimodal models. OpenTryOn integrates two variants via a single adapter, KimiUnderstandAdapter, for image and video understanding:

  • Kimi K2.6 (kimi-k2.6): General-purpose multimodal model with optional "thinking" mode and a 256K context window.
  • Kimi K2.7 Code (kimi-k2.7-code / kimi-k2.7-code-highspeed): Coding-focused variant built on K2.6 with the same vision/video understanding, tuned for long-horizon agentic coding and tool use. Thinking mode is always on for this variant.

Unlike most other adapters in this repo, Kimi's understanding capability is general-purpose -- useful for describing garments, outfits, and lookbook/runway videos in the fashion domain, but equally capable on documents, UI screenshots, product photography, or any other visual content.

For local/GPU deployment instead of the hosted API, see the open-weight Kimi-VL local model.

Prerequisites​

  1. Moonshot AI Account: Sign up at platform.kimi.ai
  2. API Key: Get one from the API Keys console
  3. Environment Variable: Set MOONSHOT_API_KEY in your .env file
MOONSHOT_API_KEY=your_moonshot_api_key

Installation​

The Kimi adapter reuses the openai package (already a core dependency of opentryon, since the Kimi API is fully OpenAI-SDK compatible). No additional installation is required.

Quick Start​

Image Understanding​

from tryon.api import KimiUnderstandAdapter

adapter = KimiUnderstandAdapter()  # uses kimi-k2.6 by default

result = adapter.understand_image(
    "garment.jpg",
    prompt="Describe this outfit: color, pattern, style, fit, and material.",
)
print(result["text"])

Video Understanding​

result = adapter.understand_video(
    "runway_clip.mp4",
    prompt="Summarize the styling and garments shown in this video.",
)
print(result["text"])

Kimi K2.7 Code (coding-focused, still multimodal)​

adapter = KimiUnderstandAdapter(model="kimi-k2.7-code")

result = adapter.understand_image(
    "ui_mockup.png",
    prompt="Write the HTML/CSS for this design.",
)
print(result["text"])

API Reference​

KimiUnderstandAdapter​

__init__(api_key=None, model="kimi-k2.6", base_url=None)​

Parameters:

  • api_key (str, optional): Moonshot API key. Defaults to MOONSHOT_API_KEY environment variable.
  • model (str, optional): Default model for calls that don't override it. One of kimi-k2.6, kimi-k2.7-code, kimi-k2.7-code-highspeed, kimi-k2.5. Default: "kimi-k2.6".
  • base_url (str, optional): Defaults to KIMI_BASE_URL env var or https://api.moonshot.ai/v1.

Raises:

  • ValueError: If the API key is missing.
  • ImportError: If the openai package isn't installed.

understand_image(image, prompt=..., model=None, thinking=None, max_tokens=None)​

Understand one or more images.

Parameters:

  • image: A single image or list of images. Each may be a file path, URL, PIL.Image, raw bytes, or BytesIO. Supported formats: png, jpeg, webp, gif.
  • prompt (str): Question/instruction about the image(s).
  • model (str, optional): Override the default model for this call.
  • thinking (bool, optional): Force-enable/disable thinking mode. Only kimi-k2.6 supports disabling it; kimi-k2.7-code* always thinks.
  • max_tokens (int, optional): Max output tokens (server default: 32768).

Returns: dict with keys text, reasoning, model, usage.

understand_video(video, prompt=..., model=None, thinking=None, max_tokens=None, use_file_upload=None, max_inline_mb=20.0)​

Understand video content.

Parameters:

  • video: File path, URL, raw bytes, or BytesIO. Supported formats: mp4, mpeg, mov, avi, x-flv, mpg, webm, wmv, 3gpp.
  • use_file_upload (bool, optional): Upload to Moonshot storage (ms:// reference) instead of inlining as base64. Defaults to auto-enabled above max_inline_mb.
  • Other parameters same as understand_image.

Returns: dict with keys text, reasoning, model, usage.

understand(image=None, video=None, prompt=..., model=None, thinking=None, max_tokens=None)​

Single entry point that accepts image and/or video (at least one required) -- this is what the opentryon understand --model kimi-k2.6 CLI command calls.

chat(messages, model=None, thinking=None, max_tokens=None, tools=None, tool_choice=None)​

Escape hatch for full multi-turn conversations or tool-calling agents (e.g. Kimi's "watch a video clip" tool-use pattern). Returns the raw OpenAI SDK ChatCompletion object.

Parameter Notes​

Kimi's k2.5/k2.6/k2.7-code models fix temperature, top_p, n, presence_penalty, and frequency_penalty server-side -- non-default values raise an API error, so this adapter doesn't expose those knobs. Only thinking and max_tokens are configurable.

FieldBehavior
thinkingDefault enabled. kimi-k2.7-code* cannot disable it.
max_tokensDefault 32768.
temperature / top_p / n / presence_penalty / frequency_penaltyFixed by the API; not exposed.

Using the opentryon CLI​

# Kimi K2.6 -- image understanding
opentryon understand --model kimi-k2.6 --image garment.jpg \
  --prompt "Describe this outfit."

# Kimi K2.6 -- video understanding, disable thinking mode
opentryon understand --model kimi-k2.6 --video runway_clip.mp4 --no-thinking

# Kimi K2.7 Code -- coding-focused multimodal understanding
opentryon understand --model kimi-k2.7-code --image ui_mockup.png \
  --prompt "Write the HTML/CSS for this design."

# High-speed K2.7 Code variant
opentryon understand --model kimi-k2.7-code --kimi-model kimi-k2.7-code-highspeed \
  --image garment.jpg

Results are printed to stdout and saved as JSON under outputs/.

Error Handling​

The adapter raises ValueError for:

  • Missing API key
  • Neither image nor video provided to understand()
  • Invalid model
  • Attempting to disable thinking on kimi-k2.7-code*

References​