ComfyUI Node
ThinkingLLM Advanced
A ComfyUI node in ThinkingLLM with 20 inputs and 2 outputs.
ThinkingLLM Advanced
- image
- video
- RESPONSE
- RAW_TRACE
◄model_nameQwen3-VL-4B-Instruct-Abliterated [DL: 7.5GB, VRAM: 6.0GB]►
◄attention_modeauto►
◄use_torch_compilefalse►
◄deviceauto►
◄preset_prompt🖼️ Detailed Description►
◄custom_prompt►
◄max_tokens8192►
◄temperature0.60►
◄top_p0.90►
◄num_beams1►
◄repetition_penalty1.00►
◄frame_count16►
◄keep_model_loadedfalse►
◄seed1►
◄keep_last_promptfalse►
◄stream_tokens_to_terminalfalse►
◄enable_thinkingtrue►
◄hf_token►
CategoryThinkingLLM
Inputs (20)
| Name | Type | Default | Description |
|---|---|---|---|
| model_name | COMBO | Qwen3-VL-4B-Instruct-Abliterated [DL: 7.5GB, VRAM: 6.0GB] | Pick the Qwen-VL checkpoint. First run downloads weights into models/LLM/Qwen-VL, so leave disk space. |
| attention_mode | COMBO | auto | auto tries SageAttention → FlashAttention 2 → SDPA in order. SDPA is stable and recommended. Only override when debugging attention backends. |
| use_torch_compile | BOOLEAN | false | Enable torch.compile('reduce-overhead') on supported CUDA/Torch 2.1+ builds for extra throughput after the first compile. |
| device | COMBO | auto | Choose where to run the model: auto, cpu, mps, or cuda:x for multi-GPU systems. |
| preset_prompt | COMBO | 🖼️ Detailed Description | Built-in instruction describing how Qwen-VL should analyze the media input. Select 'No preset' to use only the custom prompt or image input. |
| custom_prompt | STRING | Additional user input that gets combined with the preset template. Leave empty to use only the template. | |
| max_tokens | INT | 819264–8192 | Maximum number of new tokens to decode. Larger values yield longer answers but consume more time and memory. |
| temperature | FLOAT | 0.600.1–1 | Sampling randomness when num_beams == 1. 0.2–0.4 is focused, 0.7+ is creative. |
| top_p | FLOAT | 0.900–1 | Nucleus sampling cutoff when num_beams == 1. Lower values keep only top tokens; 0.9–0.95 allows more variety. |
| num_beams | INT | 11–8 | Beam-search width. Values >1 disable temperature/top_p and trade speed for more stable answers. |
| repetition_penalty | FLOAT | 1.000.5–2 | Values >1 (e.g., 1.1–1.3) penalize repeated phrases; 1.0 leaves logits untouched. |
| frame_count | INT | 161–64 | Number of frames extracted from video inputs before prompting Qwen-VL. More frames provide context but cost time. |
| keep_model_loaded | BOOLEAN | false | Keeps the model resident in VRAM/RAM after the run so the next prompt skips loading. |
| seed | INT | 11–4294967295 | Seed controlling sampling and frame picking; reuse it to reproduce results. 💡 Cache Info: Prompts are cached automatically. Use same inputs (model, preset, custom prompt, image/video) to reuse cached prompts and avoid regeneration. 🔒 Fixed Seed Mode: Set seed = 1 to ignore image/video changes and only use text-based caching. Perfect for keeping the same prompt regardless of media input variations. |
| keep_last_prompt | BOOLEAN | false | Keep last generated prompt instead of creating a new one |
| stream_tokens_to_terminal | BOOLEAN | false | Print every generated token live to the ComfyUI terminal/console |
| enable_thinking | BOOLEAN | true | Enable model reasoning/thinking when the backend supports it: True=allow thinking, False=force direct answer. Even when enabled, easy prompts may still get a direct answer, and this node automatically disables thinking when there is not enough output budget left for useful reasoning. For non-Qwen models (Gemma, LLaMA) this is advisory. |
| hf_token | STRING | Optional Hugging Face access token for private or gated model downloads. It is passed only to the download call, never logged or cached, and the in-memory copy is dropped after the download attempt. Clear this field before saving or sharing workflows. | |
| imageopt | IMAGE | — | |
| videoopt | IMAGE | — |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| RESPONSE | STRING | — |
| RAW_TRACE | STRING | — |