ThinkingLLM Advanced

If the plain ThinkingLLM node is the one you start with, Advanced is the one you graduate to. Same engine, same model catalog, same thinking toggle and RAW_TRACE output - but every sampling dial is exposed, and you get a few knobs the base node doesn't have at all. Reach for it when the default answer isn't quite what you want and you need to push the model in a direction.

It's the Transformers/HF path of the pack, so it's the easiest cross-platform option on Windows and Linux. No llama.cpp backend to babysit - just a model, a prompt, and a pile of sampling settings.

The dials that actually matter

• temperature / top_p - the classic randomness pair. The tooltip's framing is right: 0.2–0.4 is focused, 0.7+ is creative. Defaults are a sane 0.6 / 0.9.
• num_beams - beam search, 1–8. Here's the catch: values above 1 disable temperature/top_p entirely and trade speed for a more stable answer. If you've set a lovely creative temperature and the model feels stiff, check that beam count.
• repetition_penalty - nudge it to 1.1–1.3 when Qwen starts echo-chambering on long video descriptions.
• frame_count - how many frames get pulled from a video input before prompting. More frames = better context, slower run, and a bigger token bill. 16 is the default for a reason.
• attention_mode - auto tries SageAttention → FlashAttention 2 → SDPA in order. SDPA is the stable fallback and is honestly fine; only touch this when debugging attention backends.
• use_torch_compile - torch.compile('reduce-overhead') for extra throughput. The first run pays a compile cost; after that it can be meaningfully faster on supported CUDA/Torch 2.1+ builds. It's off by default, and leaving it off is a perfectly reasonable choice on a 20-series card.
• device - auto/cpu/mps, plus per-GPU targets (cuda:0, etc.) for multi-GPU rigs.

Everything else carries over from the base node: model_name, the preset/custom prompt pair, max_tokens (capped at 8192 on this path), enable_thinking, keep_model_loaded, seed, keep_last_prompt, stream_tokens_to_terminal, and hf_token for gated downloads.

Installing

Same as the rest of the pack - ComfyUI Manager, search ThinkingLLM, install. Manually:

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

Restart, and the node appears under the ThinkingLLM category. The HF path needs transformers>=5.2.0; the pack auto-installs missing tokenizer backends (sentencepiece, tiktoken) on first load.

Gotchas

First run downloads weights into models/LLM/Qwen-VL - the default 4B VL model is a ~7.5 GB pull, so it looks frozen while it downloads. The fixed-seed prompt cache is still active here: seed 1 ignores media changes, and identical inputs reuse the saved prompt, so change the seed or inputs deliberately when you want a fresh answer. And remember the beam-search/temperature conflict from above - it's the one knob pairing that silently overrides your other settings.