Local Multimodal Chat · Tank

Local Multimodal Chat · Tank is the node that does the actual work: you give it a prompt, optionally an image, a batch of images, or video frames, and a local Qwen3.8 model answers in text. That's it - no chat window, no history, no server. Wire the sampler's output image in and get back a description; wire the description into a prompt box and you have img2img seeding that runs entirely on your own machine.

This is the LLM-in-the-graph pattern: an 8B-to-27B local VLM is uncensored, offline and free per call, which an API fundamentally is not. The two things it's genuinely good for are building caption files for a training set and turning a rough idea (or a reference image) into a structured prompt you can hand to an encoder. It is not a writer, and every VLM here shares one weakness worth knowing before you trust a batch: multi-subject attribution. Two people in frame and it will mix up who's wearing what, which is why people auto-caption the large sets and audit the small ones by hand.

What you set

backend comes from Qwen3.8 Model Loader · Tank - that's the only way in. Then:

• prompt - your instruction. The default asks for a detailed description, which is a fine starting point for captioning.
• system_prompt - persona and format rules. If you want output you can pipe into another node, say so here, in plain words ("no preamble, one paragraph, no markdown").
• image / video_frames / video - optional inputs. image accepts one image or a batch and sends every frame of it, which is how you caption a folder in one queue. video_frames takes a decoded IMAGE batch and samples it down to max_video_frames (8 by default); video takes a ComfyUI VIDEO object. Local models always sample frames - video_transport exists for workflow compatibility and is ignored here.
• max_tokens (4096) and context_length (8192) - the two numbers that matter on real prompts. Context covers your prompt plus media tokens plus the answer, and images eat a lot of it; bump it in 1024 steps when output gets clipped. Note the reload behaviour: changing context reloads the model, changing reasoning effort doesn't.

Sampling is the usual llama.cpp stack - temperature 0.6, top_p 0.95, top_k 40, min_p 0.05, repeat_penalty 1.05 - and the defaults are fine. Turn temperature down for captions you intend to reuse; 0 gets you deterministic greedy sampling.

thinking_mode is the one with a gotcha: auto on a local model means thinking off, preserving the pack's original behaviour. Qwen3.8 natively supports low/medium/xhigh, so high resolves to medium and max to xhigh; the stats output shows what actually took effect. And unload_after frees the VRAM the moment generation finishes - great when the graph continues into a diffusion model, wasteful when you're iterating on the same image.

What comes out

response is the clean answer with hidden reasoning tags stripped - this is the one you wire into a prompt node or a text-save node. reasoning holds the thinking when the model exposes it, raw_response is the unedited assistant content (including a full tool-call message), and stats gives you timing, token counts, speed and backend details - check it once per session to confirm whether reasoning was on when you expected it off. The optional tools_json field takes an OpenAI function/tool schema if you're doing something more exotic than prompting.

Install

requirements.txt installs only diskcache; the real prerequisite is a CUDA build of llama-cpp-python that knows Qwen3.8's chat template. Do not let anything install the plain PyPI wheel over it. If the nodes won't load with No module named 'llama_cpp_cuda' or the pack complains about a missing Qwen35ChatHandler, that's your build, not the pack. Full setup lives in the Qwen3.8 Model Loader article; the short version is ComfyUI Manager → search Tank, or clone into ComfyUI/custom_nodes and restart, then put the GGUF and the matching projector in ComfyUI/models/LLM/Qwen3.8/.

Where people get burned

First run looks frozen. Weights load when chat starts, not when the loader runs - that's a 27B GGUF plus a projector, so give it time and watch the console, not the canvas.

Truncated answers. Almost always context_length, not max_tokens. Images and video frames consume context before the model writes a word.

Garbage answers about the image. A projector from a different release than the main model. The loader's backend_info output names the recommended projector and whether your filename matched - read it.

Reasoning text leaking into your prompt. If you're feeding the output into an encoder, take response, not raw_response.

The model says something confidently wrong about a two-person photo. That's not a bug you can tune away; it's the known ceiling of every captioner in this family.