LLM Chat (llama.cpp, integrated)

MiniMaxLLMChat is the brain of the MiniMax Music Production Toolkit: a self-contained llama.cpp chat node that turns the structured brief from the Song Prompt node into the full [Caption] / [Lyrics] / [Title] / [Image_Prompt] package that MiniMax Music 3 and your FLUX cover branch actually consume.

And here's the part that surprises people given the "MiniMax" in the name: it never calls a MiniMax API and needs no key. It's a local GGUF model running through llama-cpp-python, sitting entirely inside your ComfyUI graph. That's the same local-LLM pattern the image community settled on for prompt enhancement - a small obedient model on your own card is plenty for structured rewriting, and it costs nothing per call.

The pack used to punt this job to the external ComfyUI-LLM-Session nodes. Since v2.0.0 the chat is integrated here, so the example workflow is toolkit-plus-ComfyUI-core only.

How it works

The node loads a llama.cpp GGUF from models/llm, sends one system+user turn, and returns the assistant text. By default (reset_session ON) every run is a fresh single-turn chat - that's the recommended setting. Turn it OFF and llama.cpp session state is kept per session_id, so you can hold a multi-turn conversation. A new session_id value also busts ComfyUI's output cache, which is how you force a fresh creative pass on identical genre prompts in a batch.

Reasoning models are handled properly: with thinking set to off, the node asks the backend to skip reasoning and always splits any <think> blocks out of the answer, logging them separately. Your parser downstream only ever sees the clean answer - that matters, because leaked reasoning text is exactly the kind of junk that used to pollute generated prompts.

Inputs that matter

• enabled - master switch. Set it false and the node returns empty text without loading any model, and the rest of the workflow falls back to manual fields. Great for testing without an LLM on the card.
• model - any GGUF in models/llm. The example name Qwen3.8-27B-UD-IQ3_XXS.gguf is always offered, but that file is not bundled - you provide it or configure a download URL.
• max_tokens (example 16384) and n_ctx (example 32768) - the generous defaults exist because one response has to fit a detailed caption, a long instrumental structure map, lyrics, a title and a cover prompt. Don't shrink n_ctx casually; the production system prompt eats a real chunk of it.
• chat_format - auto picks the verified template per model family (chatml for Qwen-style models, the model's own embedded template for Gemma). The author verified Qwen3.8-27B and Gemma 4 end to end; none falls back to the GGUF's own template.
• temperature / top_p / top_k / min_p - the LM Studio-style sampling set, defaults 0.7 / 0.8 / 40 / 0.
• n_gpu_layers (-1 = offload everything), seed (-1 = random each run), and the multi-GPU quartet split_mode, tensor_split (try even or fractions like 2,3), main_gpu, tensor_parallel - only relevant with more than one GPU.

Outputs

• text - the assistant response; wire it into the pack's parser node (MiniMaxParseExternalLLMOutputV16), which splits out Caption/Lyrics/Title/Image_Prompt.
• status - a one-line report (model, session mode, character count) for diagnostics and the production JSON.
• thinking - the split-off reasoning, if any.

Installing

Everything is in requirements.txt except one package. After cloning the pack you need llama-cpp-python in the same Python that runs ComfyUI:

python -m pip install llama-cpp-python

Then drop a GGUF in ComfyUI/models/llm. If llama-cpp-python is missing the node still registers - you only get a clear error at execution time, which is polite but means the failure happens mid-run, not at startup.

Where people get burned

The model file is the #1 stumble: the dropdown shows the example filename whether or not the file exists, and with auto_download ON it will only fetch a GGUF whose URL is actually configured in models_config.json. No URL configured means a clear error naming the expected location - read it before blaming the node. And budget your VRAM: a 27B-quantized GGUF shares the card with the MiniMax generation stage that runs next. That's not this node's job to solve - the integrated unload node exists for exactly that.