AI News: 3D, Video, and Language Models for Your Own Hardware
Short news notes about what has changed in the tools: a new version has been released, the license has changed, support for another platform has been added. No press release retelling: if a claim cannot be tied to a repository, model page, or official page, there will be no note.
Articles
184 articles
LichtFeld-Studio Releases RoMa v1 Weights for Image MatchingThe LichtFeld-Studio tool has a new release — RoMa v1 dense matcher weights. This is a permanent model asset in the native .lfw format, used in the de…
Xing4.0-29B-A4B: China Telecom's new 31B-parameter MoE modelOn September 16, the model XingChen-AGI/Xing4.0-29B-A4B was published on HuggingFace — a text-generation model from the Xing series (formerly TeleChat…- PhysStream Introduces Streaming Video Generation with Physics-Based Motion ControlA group of researchers has published the PhysStream paper on arXiv — an autoregressive image-to-video model that lets you control object motion during…
MoE-JEPA boosts V-JEPA with a mixture of experts for synthetic image detectionResearchers presented MoE-JEPA, an architecture for detecting generated and manipulated images based on V-JEPA 2 with an added Residual Mixture-of-Exp…
NVIDIA Releases c-foundationstereo-s Depth Estimation Model on HuggingFaceNVIDIA published the c-foundationstereo-s model on HuggingFace for the depth-estimation task. This is FoundationStereo — a model for depth estimation…
llama.cpp v0.4.1: support for Maple 20B-A1B, Tencent Hy 4 and Spark2.5The llama.cpp v0.4.1 release came out on September 14, 2026. The update adds support for three new architectures: Maple 20B-A1B — a ternary MoE for CP…
VC-Attention Speeds Up Low-Bit Attention for Video Generators on GPUThe authors of VC-Attention: Value Smoothing and Softmax Casting for Low-bit Attention propose a training-free low-bit attention method for Diffusion…
Cheap Chinese Coding Models: What to Choose in September 2026A developer of a web project asked a simple question: which inexpensive model to hook up for coding with pay-per-token billing that also performs well…
BubbleID: Image Segmentation Model for Boiling Analysis Released on HuggingFaceOn September 14, 2026, the UARK-NED3/BubbleID model with the image-segmentation task was published on HuggingFace. It is a research package for analyz…
OOOSplat v0.4.1 improves error handling and preview stabilityOOOSplat v0.4.1 is out — an update to the local 3D Gaussian Splatting generator focused on stability and failure handling. The GitHub release describe…
Agnes-AI releases open preview model Agnes-3.0-Flash with 33 billion parametersThe model Agnes-AI/Agnes-3.0-Flash with the text-generation task has been published on HuggingFace. This is an open preview checkpoint of Agnes 3.0 Fl…
whisper.cpp v1.9.4: Windows on ARM support and ggml update to 0.23.0Release v1.9.4 of the whisper.cpp runner is out. Changes published in the GitHub repository: Windows on ARM support added to the build pipeline, the g…
Tri-DehazeGS Separates Scene and Haze in 3D Reconstruction from PhotosA group of researchers presented Tri-DehazeGS, a method for recovering clean 3D scenes from smoky or hazy images. The work was published on September…
ComfyUI v0.35.0: WAN3-Prime, Omni 1.1 support and new 3D nodesComfyUI v0.35.0 was released on 09.09.2026. The release brings support for new partner models, 3D nodes for working with meshes, and several fixes rel…
YuE2 released with open source and weights under CC BY-NC 4.0The YuE tool has a new release, YuE2 · frontier music generation with symbolic planning. Release date — 09.09.2026. The GitHub release describes three…
YuE2-3B: An Open Model for Music Generation with Editable ScoreThe m-a-p/YuE2-3B model with the text-to-audio task has been published on HuggingFace. The author is m-a-p, whose tools are already covered in our ref…
Programmable World Model — a new framework for controllable video worlds with persistent stateA paper on Programmable World Model has been published on arXiv, in which the authors propose a framework for creating interactive video worlds with a…
Test-time guidance improves the geometry of image-to-3D models from partial observationsResearchers presented a training-free method that allows partial geometric observations of an object to be taken into account at the generation stage…
Streaming talking head: authors decoupled audio and rendering to speed up generationA paper titled Decoupled Self-Forcing Distillation for Streaming Talking Head Generation has appeared on arXiv. For streaming talking head generation…
Edge0 Releases 35B MoE Model That Runs in 3 GB of MemoryEdge0-35B-A3B-preview has been published on HuggingFace — a 35B-class text model with a sparse MoE architecture. The authors claim it decodes at 15 to…
Spectral Transport Homeostasis method corrects temporal consistency in video withoutA group of researchers published a paper on arXiv about diagnosing and correcting temporal consistency in video generation. The authors introduce the…
Nex-AGI releases Nex-N2.5-mini, a 35B-parameter agentic modelThe model nex-agi/Nex-N2.5-mini with the text-generation task has been published on HuggingFace. It is the smaller model of the new Nex-N2.5 family, f…
MiniCPM5-2B: New 2B Model for Local Deployment Released on HuggingFaceOn September 6, 2026, the openbmb/MiniCPM5-2B model with the text-generation task was published on HuggingFace. It is the second model in the MiniCPM5…
SGLang v0.5.19: Qwen3.8 support, beam search, and DeepEP v2On September 5, 2026, SGLang v0.5.19 was released. The release includes 786 pull requests from 214 contributors. New models include Qwen3.8 (2.4T-A95B…
DSAQuant proposes stage-aware quantization for video diffusionThe authors of an arXiv paper showed that conventional quantization-aware training degrades precisely the details and sharpness of video models, while…
OctWorld presented as a generator of long consistent videos with octree-based 3D memoryOn arXiv, September 3, 2026, a paper on OctWorld was released — a framework for video generation based on diffusion with persistent 3D memory. From a…
Ling-3.0-flash-Fin: Ant Group's 127.5B-Parameter Financial Model Released on HuggingFaceOn September 3, 2026, the inclusionAI/Ling-3.0-flash-Fin model with the text-generation task was published on HuggingFace. It is the first model in th…- LichtFeld-Studio releases LPIPS v0.1 weights based on VGG16LichtFeld-Studio has released LPIPS v0.1 VGG16 weights. This is a persistent model asset in the native .lfw format used by the LPIPS metric evaluator…
Microsoft releases VibeVoice-ASR-Streaming-7B for streaming speech recognitionOn September 2, 2026, the microsoft/VibeVoice-ASR-Streaming-7B model with the automatic-speech-recognition task was published on HuggingFace. The mode…
CC-4DGS reduces dynamic Gaussian Splatting storage to 20–30 MB per sceneA paper titled CC-4DGS has been published on arXiv, proposing a method to reduce the storage footprint of dynamic 4D Gaussian Splatting scenes from te…