ThinkingLLM Gemma 4 Audio (GGUF):
This node uses a Gemma 4 audio-capable GGUF model to respond to spoken content. It is suited to tasks such as transcription and brief analysis of a recording. GGUF is a model-file format used by the llama.cpp backend; the node handles model loading and audio preparation so you can focus on the instruction and the result. It can also retain the model between runs or release it afterward, giving you a choice between faster repeated use and lower memory use.
ThinkingLLM Gemma 4 Audio (GGUF) Input Parameters:
model_name
Selects the Gemma 4 audio GGUF model. The choices come from the available model catalog entries that have the accompanying multimodal projector file needed for audio processing. The default is the first option whose name contains gemma-4-12b, if one exists; otherwise it is the first available option. If no matching model is found, the displayed fallback is (no Gemma 4 audio GGUF models found), which indicates that you need to make a compatible model available. The selected model affects the resources needed to run the node and the responses it can produce.
custom_prompt
The instruction you want the model to follow when interpreting the audio, such as “Transcribe the speech” or “Summarize the main points.” This is a multiline string whose default is the node’s built-in Gemma 4 audio prompt; its exact text is not included in the provided interface. A clear, focused instruction helps the model produce the kind of response you need. The node’s tooltip recommends short, 16 kHz mono WAV-style audio for best results and notes that a connected ComfyUI AUDIO input is converted to WAV before inference.
max_tokens
Sets the maximum number of tokens the model may generate in its response. The default is 2048, the minimum is 64, and the maximum is 32768. A smaller limit is often sufficient for a short transcription or summary and can reduce generation time; a larger limit gives a lengthy transcription more room but may take longer. Tokens are pieces of text, not a fixed number of words.
keep_model_loaded
Controls whether the model remains loaded after the node finishes. The default is False. Set it to True when you expect to process several clips with the same model, so subsequent runs can reuse it; leave it False when you want the node to release model resources after a run.
seed
Sets the generation seed, which helps make repeated runs more consistent when the model and other settings stay the same. The default is 1, the minimum is 1, and the maximum is 4294967295. Changing the seed can change the wording or content the model generates; it does not change the audio itself.
stream_tokens_to_terminal
Controls whether generated text is streamed to the terminal while the model responds. The default is False. Enable it if you want to watch generation progress in the terminal; it does not add another node output.
enable_thinking
Controls whether the model’s thinking mode is requested. The default is False. The node’s tooltip notes that Gemma 4 can reason, but transcription and short audio analysis are usually clearer with thinking disabled. Consider enabling it for an instruction that calls for more involved reasoning, while checking that the final response still meets your needs.
auto_finalization_retry
Controls whether the node requests an extra model completion if the first output is empty or contains only reasoning. The default is False, so one node execution normally makes one generation pass. Enable it if you encounter those incomplete results; the additional pass can increase processing time.
hf_token
An optional Hugging Face access token, entered as a single-line string. The default is an empty string. Supply a valid token if the selected model requires authenticated access for download; otherwise, leave it blank. Treat the token as a secret and avoid sharing workflows that expose it.
audio
An optional ComfyUI AUDIO input containing the recording to analyze. No default value or duration range is specified in the node interface. When supplied, this audio is passed to the model workflow and converted to WAV before inference. For best results, follow the node’s recommendation to use a short recording suited to 16 kHz mono WAV-style input.
audio_file_path
An optional single-line string containing a local audio file path. The default is an empty string. Use it when you want the node to read a recording from disk instead of connecting an AUDIO input. The path must identify a file the ComfyUI process can access; the provided interface does not specify how the node chooses between this path and audio if both are supplied, so use one audio source at a time.
ThinkingLLM Gemma 4 Audio (GGUF) Output Parameters:
RESPONSE
A STRING containing the node’s main response to your audio instruction. Depending on your prompt, this may be a transcription, a summary, or an answer about the recording. This is the output to use when you want the model’s answer in a downstream text-processing step. As with any model-generated description or transcription, review important details against the recording.
RAW_TRACE
A STRING containing the generation trace alongside the main response. It is useful for inspecting what the node captured from the model when the main response is missing or unexpected, particularly if thinking mode is enabled. Treat it as a diagnostic text output rather than assuming it is the final answer or a word-for-word transcription.
ThinkingLLM Gemma 4 Audio (GGUF) Usage Tips:
- For transcription, start with a direct prompt such as “Transcribe the speech in this recording.” Keep
enable_thinkingoff initially, as recommended by the node’s tooltip for transcription and short analysis. - Use short, clear recordings where possible. The node recommends 16 kHz mono WAV-style audio, and a connected ComfyUI
AUDIOinput is converted to WAV before inference. - Connect
audioor enteraudio_file_path, rather than supplying both. This avoids relying on source-selection behavior that is not specified in the provided interface. - Set
max_tokensaccording to the task: a short summary needs less space than a long transcription. If a response stops before the task is complete, try raising the limit within the allowed range. - Turn on
keep_model_loadedfor a series of clips using the same model. Leave it off when memory availability matters more than the speed of the next run. - If you receive an empty or reasoning-only result, try
auto_finalization_retry. It requests another completion only in that situation, so allow for additional processing time.
ThinkingLLM Gemma 4 Audio (GGUF) Common Errors and Solutions:
(no Gemma 4 audio GGUF models found)
- Explanation: The model selector found no catalog entry matching a Gemma 4 audio model with an accompanying multimodal projector.
- Solution: Make a compatible Gemma 4 audio GGUF model and its required projector available through the installation’s model catalog, then refresh the node’s model choices.
[QwenVL] Local GGUF model not found: <path>
- Explanation: A selected local model entry points to a GGUF file that does not exist at the specified location.
- Solution: Check that the file exists and that its path matches the selected model entry. Restore or relocate the model file as needed.
[QwenVL] Local mmproj not found: <path>
- Explanation: The selected local model’s accompanying multimodal projector file is missing. This file is needed for the audio-capable model setup.
- Solution: Obtain the projector associated with that model and place it at the expected path; do not substitute a projector from an unrelated model.
[QwenVL] GGUF model not found locally and no repo_id provided: <path>
- Explanation: The model file is not present, and its catalog entry does not provide a repository from which the node can download it.
- Solution: Install the model file locally or select a catalog entry that has a valid download source.
[QwenVL] mmproj not found locally and no repo_id provided: <path>
- Explanation: The projector file is missing, and the catalog entry provides no repository from which to download it.
- Solution: Install the matching projector locally or select a model entry with a valid download source.
Model access or download failure
- Explanation: The node may need to download model files, and access can fail if a repository is restricted, a token is missing or invalid, or the download cannot complete.
- Solution: Check the selected model’s access requirements, provide a valid
hf_tokenif needed, and confirm that the machine running ComfyUI can reach the model repository.
Audio file not found or cannot be read
- Explanation: The path in
audio_file_pathmay be incorrect, inaccessible to the ComfyUI process, or unsuitable for audio processing. The provided source does not specify one exact message for every such failure. - Solution: Verify the path and file permissions, or connect a valid ComfyUI
AUDIOinput instead. If the format causes trouble, try a short WAV recording.
Audio duration or shape validation failure
- Explanation: The audio-processing path includes duration and shape validation, so an unsuitable, malformed, or excessively long recording may be rejected. The exact limits and error text are not shown in the provided node interface.
- Solution: Try a shorter, valid recording with a straightforward mono audio format. If using a connected
AUDIOvalue, confirm that it comes from a working ComfyUI audio node.
Model loading fails or memory runs out
- Explanation: Loading the GGUF model and its projector can require substantial system memory or GPU memory. Available resources depend on the selected model and backend.
- Solution: Close other memory-intensive work, try a smaller compatible model if one is available, and leave
keep_model_loadedoff when you do not need to reuse the model immediately.
