TankNodes

TankNodes

English · 简体中文

A collection of ComfyUI nodes by Tank. Use local Qwen3.8 or Ternary Bonsai for text, images and video, or call other models through an OpenAI-compatible API.

Search for Tank in the node menu.

Nodes

| Node | Use | | --- | --- | | Qwen / Bonsai GGUF Loader · Tank | Select a main model and matching vision projector, then connect to local chat. | | Local Multimodal Chat · Tank | Enter a prompt, optionally add images or video, and get a text response. | | API Chat (Environment Key) · Tank | Read an API key from a server environment variable. | | API Chat (Direct Key) · Tank | Enter an API key in the node. Clear it before sharing the workflow. | | Unload Model · Tank | Release GPU memory. Connect the chat response to after to wait for generation to finish. | | Text Split · Tank | Split text into a list by a delimiter (\n for new lines, Chinese punctuation works as-is). Connected nodes run once per item; advanced options trim items or use a regular expression such as [,，]. |

Each run is a separate request, with no automatic conversation history. API nodes work on their own and do not need the local model loader.

Install

Search for TankNodes in ComfyUI Manager, or run from ComfyUI/custom_nodes:

git clone https://github.com/xzbdqian10nian/ComfyUI-TankNodes.git

Restart ComfyUI and refresh the browser. Use Python 3.10 or newer. Nothing else has to be installed by hand:

• Models: the loader lists the supported models. Any listed file that is not in ComfyUI/models/LLM/Qwen3.8/ yet is downloaded from Hugging Face the first time it runs, and an interrupted download resumes. Set HF_ENDPOINT to use a mirror.
• Runtime: local models run on PrismML llama.cpp's llama-server. On first use the plugin downloads a pinned official build for your OS and PyTorch CUDA version (Linux and Windows CUDA, Apple Silicon macOS), verifies its SHA-256 and extracts it into the plugin's own runtime/ folder, once.
• Existing llama-cpp-python: Qwen keeps using it when installed; otherwise it runs on the same llama-server. Bonsai always uses llama-server.

The plugin never replaces Torch, CUDA or GPU drivers. API nodes alone download nothing.

Optional environment variables, set before starting ComfyUI:

| Variable | Effect | | --- | --- | | TANKNODES_AUTO_DOWNLOAD=0 | Disable downloads; missing files raise an error with the download link. For offline or self-managed setups. | | QWEN38_MODEL_DIR | Use another model folder. | | BONSAI_LLAMA_SERVER | Use your own llama-server executable. | | TANKNODES_LOCAL_RUNTIME | llama-server or llama-cpp-python to force how Qwen runs. |

Quick start

1. Open 01_text_chat.json and keep the loader defaults, Qwen3.8-27B-UD-Q4_K_M.gguf and mmproj-BF16.gguf.
2. Edit the user prompt and run. The first run downloads about 17 GB of model files plus the runtime and shows progress in the console; later runs load directly.

The default pair comes from Unsloth. Other catalogued files are listed in models.json. Use a main model and projector from the same release. Other GGUF files you place in the model folder, including subfolders, also appear in the menu.

Ternary Bonsai for reverse prompting

Select the official Ternary-Bonsai-2-27B-PQ2_0.gguf (Prism ML) or Ternary-Bonsai-2-27B-Abliterated-v2-PQ2_0-MTP.gguf (abliterated v2) with the shared projector Ternary-Bonsai-2-27B-mmproj-Q8_0.gguf; a mismatched projector is rejected before anything downloads. The first run downloads about 8 GB of weights plus the runtime.

For a local Bonsai chat, the plugin starts a loopback server on demand without stopping ComfyUI. It uses at least 16K context, Flash Attention, Q8 KV cache, one concurrent slot and no MTP. In measured short Chinese creative tasks, MTP was slower; a filename containing MTP does not enable it. Bonsai auto disables thinking for text-only prompt polishing and uses low effort when images or video frames are attached. Select medium explicitly for harder requests and allow enough output tokens for the reasoning. A one-token empty answer gets one retry with another seed; reasoning that exhausts the output budget raises an explicit error. Keep the model loaded for repeated calls, or use unload after to free GPU memory for image generation. On one idle RTX 5090, either 16K variant used about 8.9 GiB and worked with ComfyUI still running. Single images, short text and a synthetic two-frame input were verified; full VIDEO objects and longer workflows need checks with actual material.

Images and video

• image accepts one image or a batch and sends every image in it.
• video frames accepts an IMAGE batch and samples up to max video frames.
• video accepts a VIDEO object. Local models sample frames across the whole clip.

For APIs, frames is the default. Use video_url if the provider accepts native video. auto tries to encode the video and falls back to frames if encoding fails; it does not retry a rejected API request with frames.

The eight examples cover text, single and multiple images, both video inputs, both API key options, and model release.

API setup

Enter the base URL, the provider's exact model ID, and the key source. For example, use https://api.openai.com/v1 without /chat/completions.

For an environment key, set the variable before starting ComfyUI:

export OPENAI_API_KEY='your-api-key'

Enter only OPENAI_API_KEY in the node. This mode allows OpenAI's official endpoint and loopback services by default. An administrator can allow other providers:

export COMFYUI_API_ALLOWED_HOSTS='api.openai.com,api.example.com'

Environment-key requests require HTTPS except for loopback services. Direct-key mode uses the key entered in the node.

For API nodes, max output tokens = 0 and temperature = 0 omit those settings and use the provider defaults. Local temperature 0 uses deterministic sampling. API outputs are response, usage, stats, reasoning and raw response. Missing reasoning is empty; unavailable usage or throughput is shown as n/a.

Use ComfyUI's interrupt button while waiting for an API response or receiving output. It stops the local request and closes the connection. Whether remote generation and billing stop depends on the provider.

Useful settings

• Local Qwen auto keeps thinking off. Bonsai auto uses no thinking for text and low for images/frames. off explicitly disables it. External API auto keeps the provider's default.
• For local Qwen3.8, high runs as medium and max as xhigh. Stats show the effective setting. Changing reasoning effort keeps the model loaded; changing context length reloads it.
• Keep the model loaded for repeated chats. Enable unload after if you need the GPU memory for another model.
• Nodes 2.0 can collapse advanced inputs. The classic canvas shows them together.

Development

See tests/README.md for test commands. Licensed under MIT.

Thanks to the Qwen team, llama.cpp and the community GGUF publishers.