Hugging Face announced that TorchTPU brings a native TPU backend to PyTorch, so Transformers, Diffusers, and Kernels can run directly on TPU, and models only need to specify device_map="tpu" to load.
FastVideo released a ComfyUI native packaged version of FastH3 OmniRef on Hugging Face, which can generate video and audio based on prompts and ordered image, video, and audio references.
FastVideo releases FastH3-QAD-v2, an 8-step video and audio model trained with NVFP4 weights and quantization-aware distillation (QAD), runnable on a single Blackwell GPU (RTX 5090, RTX PRO 6000) or DGX Spark.
Why it matters: The 8-step video and audio model lands on a single Blackwell GPU with NVFP4 quantized weights, and error data versus the BF16 teacher model is provided.
Black Forest Labs releases FLUX 3 Image Klein, a 5.6B-parameter open-weight model that supports text-to-image, image editing, and multi-reference generation with up to 10 reference images, and features native 4K generation, pixel-level editing, and bounding-box control.
Why it matters: An Open-Source Image Model with 5.6B Parameters Brings 4K Generation and Pixel-Level Editing Down to 8GB VRAM, Allowing Evaluation of Local Image Generation and Editing Solutions.
Black Forest Labs has released ComfyUI repackaged model files for Flux 3 Image Klein on Hugging Face. The original model repository is black-forest-labs/flux-3-image-klein.
Why it matters: The official documentation provides the file directory and multiple quantized versions of Flux 3 Image Klein in ComfyUI, making it easy for local deployers to choose as needed.
Black Forest Labs releases FLUX 3 Image Klein Prompt Upsampler, a 9B vision-language model that expands brief image requests into detailed prompts better suited to FLUX 3 Image Klein.
Why it matters: The official release open-sources a 9B prompt expansion model that can expand brief requests into complete prompts including lighting, composition, and subject details, and supports reference images and bounding boxes.
An unofficial int4 quantized version, drama-3-int4, has appeared for Fish Audio's 35B (3B active) multi-speaker TTS model Drama 3, with a download size of 22.1 GB, lower than the original 41.5 GB. The expert layers, attention, and deltanet layers are quantized to int4, while the shared expert MLP remains in bf16.
Black Forest Labs Releases FLUX 3 Image Klein, a 5.6B Open-Weight Model That Simultaneously Supports Text-to-Image, Single-Reference Image Editing, and Multi-Reference Image Combination Generation, with Qwen3-VL-4B-Instruct as the text encoder.
Why it matters: The 5.6B open-weight model unifies text-to-image, single-image editing, and multi-image combination into a single model, and provides quantized versions with VRAM usage details, making it easier for self-deployers to assess implementation costs.
Black Forest Labs has released FLUX 3 Image Klein Base on Hugging Face, an open-weight model with 5.6B parameters. Text-to-image, single-reference image editing, and multi-reference image composition share the same model, with the text encoder being Qwen3-VL-4B-Instruct.
Why it matters: The 5.6B open weights cover text-to-image, single-image editing, and multi-image composition simultaneously, and provide three tiers of VRAM usage—BF16, FP8, and NVFP4—making it easy to evaluate self-hosting costs.
Tencent ARC released Multics on Hugging Face, achieving scalable multi-person video generation through layered communication. The repository includes the complete 4-player Multics DiT checkpoint, configuration, Wan2.2 VAE, UMT5 text encoder, and tokenizer; the same checkpoint supports 6-, 8-, and 10-person generation without retraining. The weights are distributed under the Apache 2.0 license.
Why it matters: Tencent ARC open-sources the multi-person video generation model Multics, where the same weights can scale to 10 people, letting creators evaluate generation options for multi-person scenes.
Why it matters: 8-step sampling and KV cache reuse of text and reference image context let creators assess the speedup potential of their image generation and editing workflows.
Because the official Qwen-Image 2.1 prompt rewriter is a 9B model requiring about 20 GB of memory, the author used ML Intern in HuggingChat to build a 0.8B version that runs on CPU, achieves a 99.7% valid output rate, and consumes about one quarter of the teacher model's tokens, with a total compute cost of $16 including 8,797 annotations.
FastVideo released the experimental model FastH3 Trim, which removes 8 of the least impactful blocks from FastH3 V2's 50 MiniMax H3 transformer blocks and replaces each block's AdaLN timestep projection with a shared rank-16 basis, trained with 8-step DMD2 (checkpoint 300), enabling 8-step text-to-video generation with synchronized audio.
Why it matters: FastH3 Trim uses pruning plus shared AdaLN to bring 8-step audio-video generation to consumer-grade hardware, and can be compared against V2 to weigh image quality against size.
Hugging Face released the Open TTS Leaderboard, using objective metrics to evaluate open-source multilingual TTS and voice cloning models, shortening single-model evaluation time from weeks of voting to a few hours.
Kyutai fine-tuned OVIE for 50,000 steps on the DL3DV dataset and released kyutai/ovie-ft-dl3dv, which shares the same architecture as the base model with only the weights differing.
Kyutai has released OVIE-512 on Hugging Face. This is a base model from the OVIE series, retrained to 512×512 using the same recipe, trained for 250K steps with a global batch size of 256.
Kyutai has released OVIE-ft-re10k on Hugging Face. This is a checkpoint of OVIE fine-tuned for 50K steps on RealEstate10K multi-view data, accounting for only 2% of its 2-million-step pretraining budget. The architecture is identical to the base model, and it can be loaded directly by reusing the same code path.
Lightricks released IC-LoRA Alpha Gen based on LTX-2.5-22B, which takes ordinary RGB video as input and generates frame-aligned grayscale alpha mattes without green screens, masks, or prompts, and can handle semi-transparent elements such as hair, fur, smoke, fire, sheer fabric, glass, and water.
Why it matters: It generates frame-aligned alpha mattes from ordinary video without green screens or prompts, offering a new alternative for matting and compositing workflows.
Why it matters: Lightricks turns SDR-to-HDR into an IC-LoRA based on LTX-2.5 and provides a ComfyUI workflow and EXR output path, so it can be integrated directly into post-production pipelines.
Lightricks releases LTX-2.5-22b-IC-LoRA-Restore on Hugging Face, an In-Context LoRA based on a reference video that renders old footage — low-resolution, low-bitrate, tape, or black-and-white/sepia scans — into clean, color, high-definition footage, removing compression damage, flicker, dirt, and scratches.
Why it matters: Lightricks open-sources an old-film restoration LoRA, providing a two-stage workflow from 480-line scans to 4K and measured gains, which can be compared with existing restoration pipelines.
Tencent ARC released WorldCrafter-Base on Hugging Face, including base transformer weights along with a camera adapter and LoRA, for use with WorldCrafter inference code.
Tencent ARC released self-contained WorldCrafter-Fast weights on Hugging Face, supporting image-to-video, text-to-video, and an interactive image-to-video demo, with inference possible without WorldCrafter-Base files.
Why it matters: WorldCrafter-Fast has released standalone weights, with sampling step configurations for both image-to-video and text-to-video made public, making it easier for creators to evaluate integration costs.
Tencent ARC open-sourced GAE-D64-1B, a temporal DiT with about 1B parameters (hidden_size=[768, 2048], depth=[28, 6]) that operates on a 64-channel geometry-native latent space, supports 672×378, V=81, and uses metric Plücker camera conditioning with a Qwen3-0.6B text encoder.
The Hugging Face team rebuilt most of AUTOMATIC1111's functionality with Gradio Workflow, creating a single-canvas Workflow1111 composed of 11 media pipelines and 73 nodes.
A developer used TRL and OpenEnv to reproduce Surya Narreddi's idea of having a language model write JavaScript to paint watercolors via p5.brush. The reference image pool, RL environment, training scripts and model are all open source, and the whole pipeline runs on Hugging Face.
Gradio's built-in gr.Workflow describes AI pipelines as graphs of typed nodes, offering a drag-and-drop canvas where each node can run and intermediate results are visible; the same graph is also a REST API and can be deployed to Hugging Face Spaces with one click.
Why it matters: Gradio turns multi-step AI pipelines into a visual graph and automatically exposes REST interfaces, giving teams a basis to assess migration costs against their existing production workflows.
NVIDIA released the Magpie TTS Multilingual open-weight model with 364M parameters, supporting 12 languages. This update adds Modern Standard Arabic, Korean and Brazilian Portuguese, and improves multilingual synthesis quality through updated training data.
A Hugging Face blog post explains that Diffusers now natively supports Nunchaku's 4-bit diffusion inference; loading a quantized checkpoint only requires calling from_pretrained(), with no local CUDA compilation needed.
Why it matters: Nunchaku's 4-bit quantized inference is now natively integrated into Diffusers, allowing creators to evaluate the trade-off between VRAM usage and speed for image generation on consumer-grade GPUs.
The fourth installment of the PRX series reveals its data pipeline: pre-training data is composed of a mix of public and internal datasets, with VLMs used to regenerate long image descriptions, which are then converted into corpora suitable for streaming training. The team switched to the Qwen3-VL text encoder and computes text latents in real time during training, with measured throughput loss of approximately 3–4% (roughly one extra day for a 30-day training run). Images are uniformly encoded at JPEG quality 92, and initial re-encoding is nearly imperceptible in practice.
Hugging Face and Cerebras jointly demonstrated a real-time speech-to-speech pipeline, using Cerebras to accelerate Gemma 4 31B inference, Nvidia's Parakeet for speech recognition, and Alibaba's Qwen3TTS for speech synthesis, with each layer being modular and open-source replaceable.
A coding agent directly called two Hugging Face Spaces to generate images of Paris landmarks and reconstruct them as 3D Gaussian splats, ultimately deploying the result as a static Space gallery.
NVIDIA has launched the Nemotron 3 Nano Omni omnimodal understanding model, featuring a Nemotron 3 hybrid Mamba-Transformer MoE backbone paired with the C-RADIOv4-H visual encoder and the Parakeet-TDT-0.6B-v2 audio encoder, supporting 100+ page documents, long audio and video understanding, and agentic computer operation.
Overworld releases the real-time video world model Waypoint-1.5, which can generate real-time environments at up to 720p and 60 FPS on desktop hardware such as RTX 3090 to 5090, and adds a 360p tier for broader consumer-grade hardware, with Apple Silicon Mac support to follow.
Why it matters: Waypoint-1.5 brings real-time world models down to consumer-grade GPUs, with 720p and 360p tiers letting creators evaluate the feasibility of local interactive generation.
The Hugging Face blog publishes the second installment of the PRX text-to-image model training series, using ablation studies to document which training techniques actually work. The baseline is PRX-1.2B with 1.2B parameters, using pure Flow Matching, Flux VAE latent space, 256×256 resolution, a global batch size of 256, and 100k training steps, with data consisting of 1 million synthetic images generated by MidJourneyV6.
The Gradio team open-sources Daggr, a Python library that lets you chain Gradio apps, ML models, and custom functions with code and auto-generate visual canvases. It supports inspecting any node's output, re-running individual steps, configuring fallback nodes, and automatically saving workflow state and cached results. Daggr requires Python 3.10 or later, can be installed via pip, and workflows can be deployed to Hugging Face Spaces for sharing.
Chinese open-source models have nearly unanimously shifted to MoE architecture, with Kimi K2, MiniMax M2, and Qwen3 all adopting this approach, aiming for sustainable deployment at controlled costs. Open-source activity expanded to multimodality and agents starting February 2025, with StepFun's Step-Audio-R1.1 voice model outperforming closed-source models, and Tencent open-sourcing Hunyuan Video and Hunyuan 3D.
Overworld releases the real-time interactive video diffusion model Waypoint-1, controllable via text, mouse, and keyboard. Weights are now available on Hugging Face Hub, including Waypoint-1-Small, with Waypoint-1-Medium coming soon.
Why it matters: The real-time interactive video diffusion model is open-sourced, allowing evaluation of its real-world performance in interactive world generation on consumer-grade hardware.
NVIDIA releases the open-source reasoning vision-language model Cosmos Reason 2, which ranks first among open-source visual understanding models on the Physical AI Bench and Physical Reasoning leaderboards.