Black Forest Labs published FLUX.3 Image Klein, a 5.6B open-weights model that handles text-to-image generation, single- and multi-reference editing, and bounding-box layout control through one model. It uses Qwen3-VL-4B-Instruct as its text encoder, and Hugging Face Diffusers shipped same-day support with a ModularPipeline, LoRA training scripts, and FP8/NVFP4 quantized weights. Weights are on Hugging Face under black-forest-labs.
Unsloth released an update enabling local runs of DeepSeek-V4.1-Flash-1017 and Black Forest Labs' FLUX.3 Image Klein via GGUF, alongside multi-GPU FSDP2 training (alpha), one-command MCP setup for coding agents, and images/video generation in chat. The release bundles broad performance and hardware compatibility improvements across the Unsloth Desktop and Python stack.
Unsloth released an update enabling local runs of DeepSeek-V4.1-Flash-1017 and Black Forest Labs' FLUX.3 Image Klein via GGUF, alongside multi-GPU FSDP2 training (alpha), one-command MCP setup for coding agents, and images/video generation in chat. The release bundles broad performance and hardware compatibility improvements across the Unsloth Desktop and Python stack.
NVIDIA and Hugging Face announced NeMo Automodel, a tool integrated with the Diffusers library that enables memory-efficient fine-tuning of video and image diffusion models like FLUX.1-dev, Wan 2.1, and HunyuanVideo at scale. The tool provides sharding and latent caching utilities to handle large models efficiently.
NVIDIA and Hugging Face announced NeMo Automodel, a tool integrated with the Diffusers library that enables memory-efficient fine-tuning of video and image diffusion models like FLUX.1-dev, Wan 2.1, and HunyuanVideo at scale. The tool provides sharding and latent caching utilities to handle large models efficiently.
The Nunchaku inference engine, based on SVDQuant, has been integrated into Hugging Face's Diffusers library. Unlike weight-only quantization, Nunchaku uses 4-bit weights and activations (W4A4) for transformer layers, reducing memory usage and speeding up denoising. This makes high-quality diffusion models more accessible on consumer hardware.
The Nunchaku inference engine, based on SVDQuant, has been integrated into Hugging Face's Diffusers library. Unlike weight-only quantization, Nunchaku uses 4-bit weights and activations (W4A4) for transformer layers, reducing memory usage and speeding up denoising. This makes high-quality diffusion models more accessible on consumer hardware.
Alibaba's Qwen team released Qwen-Image-2.1 as open weights, a unified 7B model for both image generation and editing. It natively generates and edits RGBA transparent images, supports up to 10 reference images, and targets panoramas, infographics, and virtual try-ons. Day-0 support arrived from vLLM and ComfyUI, plus a live Hugging Spaces demo, giving practitioners an immediately usable open alternative to closed-source image models.
Alibaba's Qwen team released Qwen-Image-2.1 as open weights, a unified 7B model for both image generation and editing. It natively generates and edits RGBA transparent images, supports up to 10 reference images, and targets panoramas, infographics, and virtual try-ons. Day-0 support arrived from vLLM and ComfyUI, plus a live Hugging Spaces demo, giving practitioners an immediately usable open alternative to closed-source image models.
Microsoft released Mage-Flow, a compact 4B-parameter foundation model for native-resolution text-to-image generation and instruction-based image editing. The model family includes Mage-Flow, Mage-Flow-Turbo, and Mage-Flow-Edit, all available on Hugging Face. Despite its small size, it achieves state-of-the-art competitive performance.
Microsoft released Mage-Flow, a compact 4B-parameter foundation model for native-resolution text-to-image generation and instruction-based image editing. The model family includes Mage-Flow, Mage-Flow-Turbo, and Mage-Flow-Edit, all available on Hugging Face. Despite its small size, it achieves state-of-the-art competitive performance.