FL Qwen3 TTS Model Loader

Every node in this pack that actually makes sound needs one of these. The Model Loader is the entry point: it downloads, caches, and hands out the model connection that Custom Voice, Voice Design, Voice Clone, and Voice Clone Prompt all hang off. If you've come from the image side of ComfyUI, think of it as the checkpoint loader for a TTS stack - except instead of one model there are five variants, and which one you pick changes which generate node you're allowed to use.

A quick word on what you're loading. Alibaba's Qwen3-TTS family went open-weights in January 2026, after the API-only Flash models got teased in December with a lot of "look what it can do." The community was skeptical until the weights actually landed, then mostly came around: the ten-language support is the genuine selling point over English-first open TTS like Chatterbox or Kokoro. The 1.7B models are what this loader defaults to, with 0.6B variants as the low-VRAM option.

The variants, and why the choice matters

The model_variant dropdown is the only input you really have to get right, because each variant is a different model type and the generate nodes enforce it:

• Qwen3-TTS-12Hz-1.7B-Base - the one for voice cloning and fine-tuning. Has the speaker encoder that the clone nodes need.
• Qwen3-TTS-12Hz-1.7B-CustomVoice - the nine predefined speakers (Ryan, Vivian, Serena...) with an optional style instruction. Also available as the 0.6B lightweight.
• Qwen3-TTS-12Hz-1.7B-VoiceDesign - turns a plain-English voice description into an actual voice.

Load the wrong one and you'll get a cheerful, specific error telling you exactly what to switch to. That guard is honestly the best part of the pack - it's the difference between "why is this broken" and "use the CustomVoice model."

The inputs that matter

• model_variant - see above. This is the one that matters.
• dtype - defaults to bfloat16, which is fine on modern cards. Drop to float16 if your GPU doesn't do bf16, or float32 if you're chasing weird NaN behavior.
• attention - sdpa is the default and the safe choice. flash_attention_2 only works if you actually have flash-attn compiled in; picking it otherwise is a fast trip to an import error.
• device - the schema defaults to cpu, but on any half-decent GPU you want cuda. CPU inference works (the README says so), it's just slow enough to test your patience.
• force_reload - busts the in-memory cache and reloads from disk. Rarely needed; leave it off.

The output is a single model connection (QWEN3TTS_MODEL). Wire it into whatever generate node matches the variant.

Installing the pack

ComfyUI Manager - search "FL Qwen3 TTS" and install. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-Qwen3TTS.git
cd ComfyUI-FL-Qwen3TTS
pip install -r requirements.txt

Then restart ComfyUI. The first time you run the loader, it downloads the model from HuggingFace into ComfyUI/models/tts/Qwen3TTS/ - a 1.7B model is a few gigabytes, so the first run looks hung while it's just pulling weights. The pack bundles the official qwen_tts Python package in its own src/, so you don't need to install the model repo separately.

Where people get burned

The two real traps: picking a model variant that doesn't match your generate node (see above), and forgetting the first-use download. The README's hardware guidance is honest - 12GB+ VRAM recommended for comfort, 16GB RAM minimum, and people report roughly 3.5x real-time factor, i.e. generation is not streaming-fast. It's a "wait for it" model, not a real-time one. Fair trade for the multilingual quality, but go in with your expectations set.