New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
Researchers just open-sourced Motion 3-to-4, a framework that synthesizes coherent 4D dynamic objects from a single monocular video. By decoupling static 3D shapes from motion latent representations, it recovers per-frame vertex trajectories and can even transfer character motion into entirely new videos.
Overworld just dropped Waypoint-1, an interactive video diffusion model you can explore in real time using a keyboard and mouse. Powered by its new WorldEngine inference library, the 2.3B model hits up to 60 FPS on an RTX 5090 with near-zero control lag. Visual consistency and hallucinations are still rough around the edges, but you can already test the demo or run it locally.
LightOn just open-sourced LightOnOCR-2-1B, an Apache 2.0 end-to-end vision-language model that turns PDFs and scans into clean text without multi-stage pipelines. Despite its tiny 1B footprint, it beats Chandra-9B on OlmOCR-Bench and runs 1.73x faster than DeepSeekOCR. The release also includes bounding-box variants and a massive 16M-page training dataset.
Meta just open-sourced ShapeR, a conditional 3D generation framework that reconstructs casual video captures into individual, metrically accurate 3D meshes. By combining SLAM points with rectified flow transformers, it avoids baking an entire scene into a single uneditable blob. Why it matters: Creators and roboticists can now pull individual, real-scale 3D assets straight from phone-recorded video.
FlashLabs just open-sourced Chroma 1.0, a 4B end-to-end spoken dialogue model that directly understands audio inputs and replies with natural speech in real time. Built on Qwen2.5-Omni and Llama 3 under an Apache-2.0 license, it skips traditional TTS pipelines and supports instant voice cloning with a short audio prompt. It currently supports English and is hosted on Hugging Face.
Researchers just open-sourced CoDance, a new animation model that can animate multiple subjects from a single, potentially misaligned pose sequence. It breaks rigid spatial constraints and accurately rebinds motion to target characters using text prompts and masks. All code and model weights are now publicly available.
Google just introduced TranslateGemma, an open translation model family in 4B, 12B, and 27B sizes built on Gemma 3. Distilled from Gemini, the 12B model outperforms the earlier 27B baseline across 55 languages while keeping multimodal support to translate text inside images.
HeartMuLa just dropped as a powerful open-source music foundation model, generating full songs with clean vocals and accompaniment directly from text and lyrics. The 3B parameter version supports multilingual lyrics, remembers melodies across song sections, and matches top commercial tools in lyric accuracy. It basically proves full-song generation is no longer exclusive to closed APIs.
Black Forest Labs just launched FLUX.2 klein, delivering sub-second image generation and editing on consumer GPUs with only 13GB of VRAM. It comes in an Apache 2.0-licensed 4B version and an open-weights 9B model, available via API, local runtimes, and a free web demo.
Black Forest Labs just launched FLUX.2 [klein], a speedy image generation family claiming sub-second inference that beats rival models by over 30%. It arrives in 4B and 9B variants, with the lightweight 4B edition released under a permissive Apache 2.0 license runnable on 8.4GB of VRAM.
NVIDIA just unveiled PersonaPlex, an end-to-end conversational model that listens and speaks simultaneously with seamless interruptions and backchanneling. Unlike past full-duplex systems locked to a single voice, it can adopt any persona and vocal style using simple text and audio prompts. Why it matters: It finally bridges human-like conversational rhythm with complete personality customization.
Kyutai just open-sourced Pocket TTS, a lightweight 100M-parameter text-to-speech model built to run entirely on CPUs without GPUs. It clocks in at roughly 6x real-time on a MacBook Air M4 with a 200ms time-to-first-audio latency, bringing voice cloning and infinite text streaming straight to consumer hardware.
Developer yatharthsharma3501 just open-sourced NovaSR, a microscopic 52KB audio upsampler that turns muffled 16kHz recordings into crisp 48kHz sound. Built with under 10 Conv1D layers, it clocks in at 3,600x real-time speed on a single A100 with virtually zero compute overhead. It makes on-device speech cleanup and instant TTS enhancement dirt cheap.
ShowLab just open-sourced ShowUI-Aloha, a computer-use agent that learns desktop workflows directly from human screen recordings. By turning demonstrations into reusable semantic traces instead of memorizing pixels, it hits a 60.1% success rate on the OSWorld benchmark across Windows and macOS.
Researchers just introduced VIBE, a compact open-source image editor pairing a 2B Qwen3-VL with the 1.6B Sana 1.5 diffusion backbone. It handles 2K edits in about 4 seconds on an H100—and under 2 seconds with its CFG-distilled version—while outperforming much larger models at preserving original image details.
UniSH just introduced a feed-forward framework that reconstructs metric-scale 3D scenes and human bodies directly from standard video. Instead of slow per-scene optimization, it recovers scene point clouds, camera motion, and SMPL bodies in one single pass, even with shaky footage. The code is now available on GitHub for local setup.
Tencent just open-sourced HY-Motion 1.0, a 1B+ parameter text-to-motion model built on the Diffusion Transformer architecture and flow matching. It covers over 200 motion categories across 6 major classes and uses a complete pre-training, SFT, and RL loop to generate physically plausible 3D animations ready for typical production pipelines.
Fal just released FLUX.2 [dev] Turbo, an ultra-fast distilled adapter for Black Forest Labs' flagship image model. Using DMD2 distillation, it slashes inference from 50 steps down to 8, outputting 1024x1024 images in 6.6 seconds for just $0.008. While weights are free on Hugging Face, commercial use requires fal's API under non-commercial terms.
Liquid AI just rolled out LFM2-2.6B-Exp, an experimental checkpoint built entirely with pure reinforcement learning. Despite its 2.6B size, it outperforms competing 3B models across instruction, knowledge, and math benchmarks, even topping DeepSeek R1-0528 on IFBench despite being 263 times smaller. The weights are already live on Hugging Face.
Alibaba's Qwen team just open-sourced Qwen-Image-Layered, a model that natively decomposes flat images into isolated RGBA layers. Users can prompt it to output 3 to 10 distinct layers and drill down recursively for infinite detail. It basically gives creators Photoshop-grade editable layers straight out of the box, skipping manual cutout work.
