ThinkingLLM

ThinkingLLM started with a three-hour debugging session that wasn't a bug at all. The author watched a Qwen node sit there, progress bar dead, output empty, and assumed it had crashed. It hadn't - Qwen had silently entered its internal reasoning mode. This node is the answer to that: a local Qwen-VL / Gemma 4 interface with a thinking toggle, live token streaming, and a raw trace output so you can finally tell "the AI is thinking" from "the AI is stuck."

It's the flagship of the ComfyUI-ThinkingLLM pack, which wraps Qwen3.5, Qwen3-VL, Qwen2.5-VL, and Gemma 4 behind a cleaner node interface than the previous Qwen-VL nodes. It runs fully locally through Hugging Face Transformers - no API, no key - and it's a GPL-3.0 fork of Deaquay's Qwen3.5-Uncensored, itself a descendant of the 1038lab QwenVL lineage.

How it works

Pick a model and the node loads it into memory (first run downloads the weights into ComfyUI/models/LLM/Qwen-VL - the 4B VL models run about 7.5 GB download, so leave disk space). Feed it an optional image and/or video, give it a preset_prompt or your own custom_prompt, and it returns a cleaned RESPONSE string plus RAW_TRACE, the full generation stream including any thinking blocks.

The party trick is enable_thinking. On, and the model may reason before answering - and for easy prompts it may skip reasoning entirely, which is expected, not a bug. Off, and it sends a /no_think-style directive to force a direct answer. The toggle works best on Qwen3/Qwen3.5-style models; other architectures treat it as advisory.

The inputs that matter

• model_name - pick from the catalog (Qwen3.5, Qwen3-VL Instruct/abliterated/unredacted, Gemma 4). Tooltips show download size and VRAM per model.
• preset_prompt / custom_prompt - presets cover "Detailed Description" plus LTX/Wan video-scene templates; custom_prompt is combined with the preset.
• enable_thinking - the headline toggle.
• stream_tokens_to_terminal - print tokens live; this is the debugger for "stuck" prompts.
• keep_model_loaded - keep it resident between runs to skip reloads; turn off when you need the VRAM back.

Also worth knowing: prompts are cached. Identical inputs reuse the saved result, and seed = 1 is "fixed seed mode," which ignores image/video changes entirely - so if you swap the image and get the same old answer, that's the cache, not a hallucination. That trips up more people than anything else in this pack.

Installing

ComfyUI Manager → Install via Registry → search ThinkingLLM. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/goodguy1963/ComfyUI-ThinkingLLM.git
cd ComfyUI-ThinkingLLM
pip install -r requirements.txt

Then restart ComfyUI. The HF path needs transformers>=5.2.0 and pulls in sentencepiece and tiktoken for Gemma/Qwen tokenizers; the pack tries to auto-install a missing tokenizer backend at first load.

Common issues

First run is slow because it's downloading multi-gigabyte weights - check the terminal. Big images blow up the context window, so resize or scale before the node (the README says so explicitly). And when the node sits at zero output, resist the urge to force-quit: turn on token streaming first. That's the whole reason this pack exists.

One word of caution that applies to every node in this family: LLM-vision custom nodes execute arbitrary Python on import, and the community has a scar from the 2024 ComfyUI_LLMVISION malware incident. Install via the Manager registry, and treat any LLM node with a sketchy provenance carefully. This pack's open fork lineage is a decent signal, but you're still running a local model loader - worth knowing what that means.