Kimi Service Connector 🦉| OwlV

Why you'd reach for it

Kimi (Moonshot AI) is another China-accessible LLM option, and its whole selling point is context length: up to 128k tokens in the largest model tier. That matters more for translation than people expect. If you're translating a long storyboard script, a whole caption batch, or a big slab of dialogue in one go, an 8k-context model chokes or silently truncates. Kimi's 32k default is a comfortable middle, and 128k is the option when the text genuinely is huge.

How it works

The node builds a MoonshotConnector pointed at https://api.moonshot.cn/v1/chat/completions, authenticates with your api_key, and sends an OpenAI-style chat-completions request. The LLM Translator node drives the call; the reply is parsed from choices[0].message.content. Real API call, real billing - one of the pack's genuinely working connectors.

The inputs that matter

• api_key - required, password-masked. Your Moonshot platform key.
• model_select - moonshot-v1-8k, moonshot-v1-32k, moonshot-v1-128k, kimi-v1, Custom.
• custom_model_id - optional; used when model_select is "Custom".

Output is llm, wired into LLM Translator's llm input.

A practical note on the model dropdown: the "v1" names are Moonshot's older lineup - the pack's list is a snapshot. If Moonshot has since renamed its current models (they iterate), pick Custom and type the current id. For most translation work 32k is plenty; reach for 128k only when the source text actually approaches that - bigger context costs more per call.

Installing it

Ships in the ComfyUI-MultiTranslator pack by OwlvChirotha ("OwlV"), installed as one unit. ComfyUI Manager → search "ComfyUI-MultiTranslator" → Install → restart. (The README's "ComfyUI-Translator" search string is stale.) Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/OwlvChirotha/ComfyUI-MultiTranslator
cd ComfyUI-MultiTranslator
pip install -r requirements.txt

Restart and look under Text Processing/LLM. Dependencies are just googletrans and requests; no model files land on your disk.

Troubleshooting

• Errors and auth failures surface as the translated output text, not a crash - check the node output before assuming the workflow broke.
• The masked key still lives in the workflow JSON; scrub it before sharing.
• With 128k context, sending very long text is tempting - just remember the input length also drives the cost. Long-context models bill by the token too.