New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
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.
OpenClaw just launched version 2.0, its biggest update yet packed with over 16,000 pull requests from 933 contributors. The release rebuilds the browser app, slashes setup friction by letting you plug in existing ChatGPT or Claude subscriptions, and adds shared cloud sessions so teams can collaborate without losing agent context.
ZAI just open-sourced the full model weights for GLM-5.3 on Hugging Face, packing 753 billion parameters focused on heavy coding and complex agent workflows. It uses the same base model as GLM-5.2, squeezing every performance gain out of pure post-training to rival top commercial models. Community-quantized GGUF versions are already out for local multi-GPU setups.
Tencent just open-sourced Hy4 preview, a 770B-parameter MoE model (49B active) with a 1M context window tailored for real-world coding, finance, and scientific tasks. Crucially, the model helped train itself through a recursive self-improvement loop for data strategy and kernel optimization. It is already rolling out across Tencent products like Yuanbao and CodeBuddy.
Researchers just open-sourced Code World Model, decoupling world logic from visual rendering. Instead of merely predicting the next frame's pixels, a coding agent acts as the 'brain' to update persistent scene states in code, which a fine-tuned MiniMax-H3 video model then renders. Why it matters: It stops interactive AI worlds from forgetting objects, keeping rules and consequences coherent over long horizons.
BreezeBlue just released open weights and inference code for Breeze-TTS-2, a bilingual text-to-speech model built for real-time interaction. It delivers sub-40ms time-to-first-audio on an H100 and lets developers design distinct voices from scratch using natural language prompts or inject vocal cues like sighs and laughs directly into transcripts. The code is Apache 2.0, while model weights remain research-only.
Researchers just open-sourced FixAnything, an AI tool that repairs visual glitches across 3D representations. When novel views in 3D Gaussian Splatting, NeRF, or meshes suffer from holes and floaters, it leverages a fine-tuned Wan2.1 video diffusion model to turn degraded walkthroughs into clean, 3D-consistent clips.
4DAnyone just dropped as an open-source framework that turns a single monocular video of someone into an interactive 4D Gaussian Splatting model. By combining skeleton tracking with multi-view diffusion, it eliminates drift and maintains clean consistency across dozens of camera angles—and the entire 12GB model is free to run locally.
Zach Tratar just open-sourced OpenHistory, a Mac app that monitors your workday and turns computer activity into hourly and daily summaries. Everything stays completely on your device using local inference, and it connects directly to local AI agents via secure MCP.
ItusiAI just open-sourced Magiviz, a turnkey platform that automates video production across five steps, from scriptwriting and character design to storyboarding, scene generation, and final stitching. It connects models like Veo, Kling, and Wan, and comes preloaded with Stripe billing and user management ready for deployment.
Audio8 just dropped Audio8TS, an ultra-compact text-to-speech model with roughly 100M parameters and a tiny 1.7GB footprint. It pulls off zero-shot voice cloning from just a few seconds of reference audio and even ships with an ONNX INT8 build for low-resource CPU inference. Running high-quality voice cloning locally on a basic laptop is now genuinely practical.
Block just open-sourced Berd, a canvas-based desktop environment built to coordinate specialized AI agents directly alongside your local files and repos. Users can spin up agents with distinct skills and personalities while swapping underlying models on the fly. It is a thoughtful push to move AI interactions away from disposable chat windows toward durable project building.
Researchers just open-sourced VoiceMem, a dedicated long-term memory system built specifically for real-time voice agents. It splits memory into a left brain for facts and a right brain for emotion, speculatively fetching memories while the user is still speaking with just 134 ms latency. Why it matters: It keeps voice companions responsive without freezing to dig through conversation logs, cutting per-turn overhead to roughly 430 tokens.
yetone just open-sourced Cumora, a cross-platform team chat where AI agents work as first-class teammates alongside humans. Instead of waiting to be poked, agents hold persistent memory, claim Kanban tasks, coordinate with each other, and send real emails using cloud pods or local models. Why it matters: It moves AI from reactive chatbots to proactive office colleagues.
Alibaba's Qwen models have just crossed 3 billion global downloads over the past six months, overtaking Meta and Google as the world's leading open-weight AI family according to Hugging Face. In 2026, Google saw 418 million downloads and Meta reached 227 million. With over 460 open models and 300,000 community derivatives, Alibaba is rapidly cementing itself as the backbone of open-source AI.
Alibaba just released the open weights for Qwen3.8-27B, a native multimodal dense model under the Apache 2.0 license. At 27B parameters, it outperforms Qwen3.7-Plus and features a 262K native context window extendable to 1M tokens. Open weights for the Max-level Qwen3.8-2.4T-A95B are also available on Hugging Face and ModelScope.
