ZONOS2 Model Loader

Every ZONOS2 workflow starts here. This is the node that pulls the checkpoint down from Hugging Face, decides whether your GPU can actually hold it, and hands a zonos2_model bundle to the two generation nodes in this pack. It's the least glamorous node on the page and the one beginners trip on first, mostly because they don't realize how much is about to download.

ZONOS2 is Zyphra's open-weights TTS model - a big MoE thing trained on 6+ million hours of multilingual speech, Apache-2.0 weights. This pack (Saganaki22's Zonos2_TTS-ComfyUI, released June 2026) wraps it to run natively in ComfyUI, and the loader is where the heavy lifting lives.

What it loads, and how

The loader's model dropdown offers two presets plus anything you drop into ComfyUI/models/zonos2/ yourself:

• ZONOS2 BF16 (drbaph/ZONOS2-BF16) - the full-quality checkpoint, roughly 14.3 GiB for the main model alone.
• ZONOS2 FP8 Mixed (drbaph/ZONOS2-FP8) - ~9.78 GiB. FP8 only on the MoE expert gate/up projections, everything sensitive kept in BF16. This is a memory-saver, not a speed hack; the README says outright it's not guaranteed faster than BF16.

First run auto-downloads the main checkpoint plus two supporting pieces into ComfyUI/models/zonos2/: the dac_44khz decoder (turns the model's audio tokens into actual sound) and the speaker_encoder (only needed for voice cloning, fetched lazily). That's why the first load takes a while. Existing folders get reused, and with download_if_missing on, only the missing bits come down - flip it off for a fully offline setup.

Beyond the model choice, three settings matter:

• dtype - auto (default) preserves the checkpoint's native dtype; mixed FP8 accepts auto or bf16 but rejects fp16.
• attention - auto uses FlashAttention when it's installed and compatible, otherwise falls back to PyTorch SDPA.
• download_if_missing - described above.

The one output, zonos2_model (ZONOS2_MODEL type), wires straight into ZONOS2 Voice Generation or ZONOS2 Voice Clone. That's the whole graph.

Where the memory magic happens

This is the part ComfyUI has been moving toward for years: the model is registered as a real module with ComfyUI's model management and AIMDO, not just dumped into VRAM. If your GPU's total VRAM is below ~17.3 GiB (BF16) or ~12.8 GiB (FP8), the main model runs through AIMDO's file-backed paging - only the MoE experts that are actually being routed get paged into VRAM, and the Memory Visualization graph shows real residency. Above that, you get the static GPU path. Change model, dtype, or attention and it hard-unloads the old bundle first, so a settings change that looks like a freeze is actually the node freeing tensors.

Installing it

Via ComfyUI Manager, search ZONOS2 TTS, or install straight from the pack's repo. From ComfyUI/custom_nodes:

git clone https://github.com/Saganaki22/Zonos2_TTS-ComfyUI.git
../venv/bin/python Zonos2_TTS-ComfyUI/install.py

Then restart ComfyUI. The installer only adds lightweight deps (accelerate, huggingface-hub, safetensors, tqdm) and leaves torch/torchaudio alone. The one dependency you do have to watch: this pack needs Transformers 5.0–5.12 (tested baseline 5.3.0) and won't work on the 4.x most ComfyUI installs still pin. install.py will tell you if you're out of range, but it won't upgrade Transformers for you - that's your job.

Common issues

• FP8 reports an unsupported format. Update the pack and fully restart ComfyUI. Version 0.1.7 rejects the retired all-layer FP8 layout.
• FlashAttention unavailable. Use auto or SDPA; auto handles the fallback.
• CUDA out of memory. Switch to the FP8 preset, unload other models, or lower max_new_tokens in the generation node. BF16 plus the DAC sits around 14.7 GiB before you even start generating.

If you don't have room for the full BF16 beast, reach for the FP8 preset - it's the difference between running this and watching it OOM.