Stable Diffusion 3.5 hits Diffusers with an 8B model and a 4-step turbo variant

Stability AI’s largest open image model just landed in Hugging Face Diffusers, and it’s a chunky one. The release includes two distinct 8-billion-parameter checkpoints for Stable Diffusion 3.5: a full-precision Large model and a “Large Turbo” variant that’s been timestep-distilled for few-step inference. If you’ve been following the space, the core architecture will feel familiar to anyone who worked with SD3 Medium, but there are two critical changes inside the transformer. It now uses QK normalization, which has quietly become standard for training massive transformer models, and the MMDiT blocks swap single attention layers for dual attention layers per modality stream.

The text encoders, VAE, and noise scheduler are all unchanged from SD3 Medium. What is wildly different is the memory footprint. Because this isn’t the medium-sized model anymore, running the full 8B parameter version at torch.bfloat16 precision is going to crush most consumer GPUs without optimization. The team clearly recognizes this, as they’re pushing quantization from the jump. The guide shows how to shove the transformer into 4-bit NF4 precision using bitsandbytes, and crucially, they recommend the bnb_4bit_compute_dtype stay at torch.bfloat16. Pair that with sequential CPU offloading, and you might actually fit this thing on a single GPU without watching your VRAM evaporate.

For the speed-obsessed, the Turbo checkpoint is the headliner. By ditching classifier-free guidance and baking the distillation directly into the model, you can drop inference steps from 40 down to just 4 to 8 while keeping guidance_scale pegged at 1.0. That’s not a typo. The trade-off, which the post doesn’t explicitly detail but which has been consistent across every distilled diffusion model, will almost certainly show up in reduced prompt adherence or compositional complexity compared to the full-fat version. The model remains gated, so you’ll need to accept the license terms and authenticate with huggingface-cli login before downloading.

The implication here is pretty stark for the broader ecosystem: open-weight image generation is now truly pushing into the territory that was exclusively held by closed APIs like Midjourney v6 or DALL-E 3 just twelve months ago. The fact that you can now fine-tune an 8B-parameter model with LoRAs on quantized hardware, as the post teases, means we’re about to see a wave of highly specialized SD3.5 derivatives hitting Civitai and other community hubs before the month is out.