VAST AI Research just open-sourced Mira-Scene, a pipeline that reconstructs fully editable 3D scenes from a single 2D image. By recovering dense, pixel-aligned correspondences, it generates and places distinct 3D objects that match the source perspective perfectly, ready for Blender animation or robotics simulation. Why it matters: It beats closed-source models like SAM3D without requiring manual layout annotations.
Cua just open-sourced CUA-S1-FORMS, the first in its family of small, specialized models built for computer use. Instead of clicking through fields one step at a time, it processes entire forms in a single pass using four core actions and field values pulled from a document. The release is MIT-licensed with the model, dataset, and code all public.
prism-ml just launched Bonsai-2, a collection of radically compressed models that restrict weights to -1, 0, and +1. The approach shrinks a 27B-parameter base model to under 6GB—slashing its footprint by over 9x while preserving baseline performance on consumer-grade GPUs and Apple silicon.
Researchers just released JEPA-Anything, an open-source framework designed to model predictive dynamics across biology, physics, and medicine. Instead of monolithic guesses, it uses Orthogonal Predictive Factorization to split latent states into independent components before recombining them. The repository provides core code and task-design tools, though pre-trained weights are not included.
NetEase Youdao just open-sourced Confucius4-R2T2, a real-time streaming speech recognition model built on Qwen3-ASR. It delivers an average latency between 200 and 600 ms and uses an append-only output mode that never rewrites previously committed words. That means live captions and real-time voice agents run without annoying screen flicker or mid-sentence rollbacks.
CopilotKit just open-sourced OpenMuse, a personal AI agent you can self-host with a persistent Chromium browser, an isolated Linux terminal, and a local file workspace. It works through multi-step plans continuously and lets you take manual control whenever you want. Why it matters: If you want an autonomous assistant for emails and browsing without handing your personal data to big tech, this gives you full control.
Zhongguancun researchers just released ZGCM-1, an open 7.39B model specialized in mathematical reasoning and agentic web search with 256K context support. It hits 97.13% on MATH-500 and comes with fully disclosed data and training pipelines ready for consumer GPUs. Why it matters: Small, highly specialized open-source models are becoming genuinely capable at complex reasoning tasks without enterprise hardware.
Edge0 just open-sourced a streaming MoE inference framework that treats storage like memory, running 35B models in roughly 2.5 GB of active RAM on iPhones and Macs. By predicting routing one step ahead to hide SSD read latency, it squeezes large models onto low-memory consumer devices without crushing performance.
AI research group m-a-p just open-sourced YuE2-3B, a music model that generates songs by drafting editable sheet music first. Users and agents can tweak melodies and chords before synthesizing full tracks with vocals, matching Suno v5 and v6 benchmarks on an RTX 4090.
Tencent just open-sourced AuK, a 1.5B foundation model that lets you generate and edit speech using simple text prompts. Beyond zero-shot voice cloning, it can seamlessly insert or delete words from existing audio, tweak emotional tone, and remove background noise. It brings text-editor simplicity to voice workflows and runs on consumer GPUs.
Researchers just introduced Fire3D, an open-source feed-forward framework that turns a single photo or casual video into interactive 3D scene assets in under 60 seconds. Running on a single GPU, it automatically decomposes up to 16 objects into full textured meshes with 6-DoF poses. Why it matters: It slashes scene reconstruction from hours to seconds, creating physics-ready assets directly usable in game engines and robotics sims.
OpenBMB just open-sourced MiniCPM5-2B, a compact 2-billion-parameter dense model that punches above 4B-class models like Qwen3.5-4B across coding and math benchmarks. Weighing only around 5GB, it easily runs locally on consumer edge hardware, and the team also released the full high-quality dataset behind it.
Google just open-sourced ARTEMIS, an automation framework that lets coding assistants like Claude Code, Cursor, and Codex drive real Android devices or emulators via MCP. It runs cross-app workflows, grabs logcat crashes and screenshots, and scores a 99%+ success rate on AndroidWorld.
Alaya Lab and the University of Tokyo just released WorldSculpt, an open-source framework that reconstructs cluttered scenes into hundreds of separate 3D meshes. Instead of fusing everything into one uneditable lump, it lets you isolate, move, and edit every single object, and even converts Gaussian splat worlds into meshes.
Researchers just open-sourced UniMate, a foundation model that animates virtually any rigged 3D asset from a single text prompt. Unlike previous animators locked to human rigs, its topology-aware diffusion transformer drives birds, serpents, and mechanical objects zero-shot with no per-skeleton retraining. The full codebase, training pipeline, and weights are publicly available.
inclusionAI just open-sourced LLaDA-Image, a 6B unified diffusion model family that handles both text-to-image and instruction-guided editing from a single checkpoint. It delivers photorealistic rendering alongside bilingual Chinese-English text, while a 4-step distilled Turbo version enables near-instant generation. Checkpoints and ComfyUI integrations are already live.
Researchers just open-sourced SolarWM, a world model framework that turns video generators into real-time interactive worlds lasting over an hour. Trained on five-second clips, it keeps environments coherent across long rollouts and fully opens a 25 TB dataset alongside recipes for models like Wan and Minimax H3.
H3 World just turned the open-source Miniax H3 video generator into an interactive game engine. Instead of building a complex control rig, it simply translates keystrokes into short text prompts that guide the video's motion in real time. The project code, local setup instructions, and training scripts are now fully open-sourced.
Shanghai AI Laboratory just open-sourced InternLumina-U2, a 16.9B diffusion MoE model with only ~1.4B active parameters. It combines text QA, image generation, editing, video, and 3D understanding into a single framework using an 8-codebook discrete visual tokenizer.
Anthropic just released commerce-agents, an open-source blueprint for deploying Claude in e-commerce. It features customer-facing shopping agents alongside back-office merchant agents across retail, travel, telecom, and entertainment verticals. Crucially, human approval gates hold all writes, ensuring agents never place live orders or charge credit cards directly.
Perplexity just launched PII-TRACE and PII-Tracer, a benchmark and 0.6B local model built on Qwen3 to screen sensitive data before routing tasks to cloud models. It achieved the highest character F1 among 12 detectors and stayed accurate across multi-turn chats. Why it matters: hybrid AI only works if on-device gates know exactly what to hide.
Google Research just open-sourced TimesFM 3.0, its third-generation time-series foundation model that now natively handles multivariate forecasting. It evaluates multiple correlated trends and planned future events with zero-shot accuracy, runs on just 1GB of VRAM, and completes full-horizon forecasts in a single forward pass under a non-commercial license.
Viggle just open-sourced Viggle-Animate, an AI pipeline built on MiniMax-H3 that ditches traditional pose skeletons, masks, and face trackers. You simply edit one frame of a video to introduce the new character, and the model propagates it across the entire shot in three forward passes, rendering 124 frames in 26 seconds on a single B200 GPU.
