Comfy-Org Makes Black Forest Labs' FLUX.2 Dev Ready to Run Locally
Black Forest Labs' 32B open-weight image model lands in ComfyUI with FP8 weights, a 24B Mistral text encoder, and 10-image multi-reference editing.
- Comfy-Org/Flux-2-Dev packages FLUX.2 [dev] weights for drop-in ComfyUI use, now past 1M downloads.
- 32B total parameters: 12B rectified flow transformer plus a 24B Mistral-3 vision-language text encoder.
- New 32-channel VAE (vs 16 in FLUX.1) released separately under Apache 2.0 license.
- Multi-reference conditioning up to 10 images, 4MP output, and sharply improved in-image typography.
- FP8-mixed quantization from NVIDIA cuts VRAM by 40% and boosts throughput 40% on RTX GPUs.
- Open weights are non-commercial by default; commercial use requires a license from Black Forest Labs.
Comfy-Org has published a pre-quantized build of Black Forest Labs’ FLUX.2 [dev] in a ComfyUI repository that showed more than one million downloads at publication. Hugging Face’s counter includes repeat file requests, so the figure measures repository activity rather than unique installations.
The package removes much of the usual setup work. Its files map directly to ComfyUI’s model folders, and the included workflows run without a Diffusers wrapper. Developers still face substantial hardware requirements: FLUX.2 [dev] combines a 32-billion-parameter rectified-flow transformer with a 24-billion-parameter Mistral vision-language encoder.
The Bundle Mirrors ComfyUI
The repository organizes its weights under split_files/diffusion_models, split_files/text_encoders, split_files/vae, and split_files/loras. Each directory corresponds to a folder under ComfyUI/models.
| Component | Artifact | Size and options |
|---|---|---|
| Diffusion transformer | flux2_dev_fp8mixed.safetensors |
35.5 GB file |
| Text encoder | mistral_3_small_flux2 |
BF16, FP8-mixed, and FP4-mixed variants; the 18 GB FP8 file offers a practical middle ground |
| VAE | flux2-vae.safetensors |
336 MB, with 32 latent channels |
| Optional acceleration | Two ByteZSzn Turbo LoRAs | Community adapters intended to reduce sampling time |
A typical installation follows five steps:
- Update ComfyUI to a build that supports FLUX.2.
- Copy the transformer into
ComfyUI/models/diffusion_models. - Choose one Mistral encoder variant and place it in
ComfyUI/models/text_encoders. - Copy the VAE into
ComfyUI/models/vae, then add any optional adapters underComfyUI/models/loras. - Import one of the supplied workflow JSON files for text-to-image generation, image editing, or product mockups.
Loader errors involving an unknown architecture or missing nodes usually indicate an outdated ComfyUI installation or custom node. Users seeking a smaller transformer can try
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
