Gemini3-LLM

Most "LLM in ComfyUI" nodes are narrow - a prompt enhancer, a captioner, a tagger. This one is the opposite: it's a full Gemini 3 chat model wearing a node costume. Text in, text out, with system prompts, web search, structured output, and the ability to look at images, video, and files you wire into it. It's the pack's general-purpose brain - you can use it to rewrite a rough idea into a structured prompt, to describe an image so another node can use the description, or to run a genuinely multi-modal analysis pass. If you've only seen the tokenizer-LLM view of Gemini, this is the "drop it in the graph and ask it anything" version.

How it works

Your prompt (and optionally image_input, video, or a file from the pack's FileLoaderNode) is POSTed to the gateway with the model id selected from the model dropdown - Gemini 3 Pro Preview, Gemini 3 Flash Preview, or Gemini 3.1 Pro Preview. The response comes back as a STRING (the output port), and a context port carries the conversation state, exactly like the image nodes. That's the part that makes this more than a single-shot API call: feed context into the pack's ContextNode, or chain it into another Gemini node, and you get real back-and-forth. The README frames it as "connect context management and it reads the context directly, or short-circuit multiple LLM nodes together."

The inputs that matter

• model (enum) - Pro vs Flash is the classic quality/speed/cost trade. Start Flash; escalate to Pro when the answer matters.
• System_prompt (STRING) - the system role. If you know how to steer an LLM, this is where you do it.
• thinking_level (Minimal/Low/Medium/High) - reasoning effort. High thinks longer and costs more.
• media_resolution (Default/Low/Medium/High) - how closely the model inspects attached media.
• Web_search (BOOLEAN) - let the model fetch current facts instead of relying on training data.
• format (BOOLEAN) - structured-output mode; turn it on when you want a predictable shape (like JSON) back out of the output string.
• seed (INT) - reproducibility.

Optional: prompt, image_input (IMAGE), video (VIDEO), file (FILE), context (ANY). Outputs: output (STRING), context (ANY).

Installing it

Part of comfyui-MJAPI-party. ComfyUI Manager → "mojieapi_party" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/MoJIeAIGC/comfyui-MJAPI-party.git

Register at mojieaigc.com and set your key in config.ini (leave BASE_URL = https://www.mojieaigc.com/v1/completions as shipped), then restart:

[API]
KEY = your_api_key
BASE_URL = https://www.mojieaigc.com/v1/completions

No local models, no GGUF files, no Ollama server - that's the entire point of this pack's LLM story.

Common issues

• The prompt shows up as a literal text widget instead of a wire. prompt is a forced-input port: it expects a wire from another node (say, a text from a prior LLM turn). Type in the widget if you like, but to build chains, wire text into it.
• Conversation "resets" between runs. Same global-context caveat as the rest of the pack: history is in-memory and only flows through the context port. Feed it forward explicitly or use ContextNode.
• Structured output isn't actually structured. format on is a request, not a guarantee - validate what comes back, and remember a chat model's JSON habit is "json plus a preamble." The pack has a JSON-parser node for exactly this, so plan for it.
• Cost. Frontier-chat quality via a reseller is metered, and thinking_level High on Pro burns through the balance noticeably. For the LLM-in-graph jobs that matter (prompt rewriting, captioning, analysis) the KB's advice holds: small and obedient beats big and clever, and here that maps to Flash + low thinking for the cheap, fast iterations.