MIT researchers found that OpenAI’s GPT-6 Astra can inspect and directly rewrite neural network weights without any gradient descent or retraining. In one test, Astra modified just 0.016% of a small Qwen model’s weights to eliminate repetitive loops, bumping its math score from 7.9% to 48.1%. If frontier models can reliably program weight space, AI self-improvement might arrive much faster than expected.
Researchers just open-sourced NitroFusion 2, slashing image diffusion down from 8 steps to just 1 or 2 steps. By introducing adversarial distillation directly into the multimodal text stream, it preserves fine details and legible typography even in a single pass. The weights are available on Hugging Face and run locally via ComfyUI.
Researchers just dropped Any Time, Any Space (ATAS), an open-source framework solving scene drift in long-form AI video. By reconstructing scenes in 3D and storing dynamic spatial memories, cameras can leave a room and return without anything changing. It plugs into models like Wan and MiniMax, with code ready to run locally.
World Play 2 just open-sourced an interactive world model that lets you roam and interact with characters and objects in real time. Instead of degrading over long horizons, it compresses past events into memory tokens to keep environments consistent even when you leave and return. It offers a 4-step Turbo version, though the full unquantized model weighs 69GB.
Alibaba just released Qwen3.8-LiveTranslate, a real-time interpretation model that cuts average latency to 2.3 seconds across 60 languages. It reconstructs speech into an interleaved audio-text stream and clones speakers' voices so translations sound just like them, making everyday bilingual meetings far less clunky.
Alibaba just rolled out Qwen 3.8 Omni Flash, an omni multimodal model featuring a 1M token context window. It handles real-time streaming video and audio inputs, automates tasks like video editing, and undercuts Gemini Flash on pricing. It is currently API-only and not open-source.
Alibaba just rolled out Qwen3.8-Max-0902, an upgraded snapshot targeting complex engineering code and collaborative agent workflows. It keeps the 1M context window and thinking mode while sharpening tool orchestration and multimodal vision, priced at $2 per million input tokens and $6 for output.
Alibaba just released Qwen-Drive-1.0, a vision-language foundation model tailored for autonomous driving. Built on Qwen3.5-4B without changing the underlying architecture, it plugs in a BEV perception head and a diffusion planning expert to handle 3D sensing and trajectory generation. It shows general VLMs can take the wheel without rebuilding them from scratch.
Alibaba just open-sourced qwen-audio-agent, a full-duplex voice runtime bridging conversational audio with background task agents. Users can converse and interrupt naturally while backend agents like OpenCode or Claude Code handle long-running code edits and report back when finished.
ABot-World just open-sourced a real-time world simulator that lets you continuously explore and steer an AI-generated environment with action controls. It runs on a single desktop RTX 5090 GPU, delivering 720p video at 16 FPS with 1.2-second latency. Basically, instead of short video clips, you get an endlessly unfolding world running entirely on local consumer hardware.
Alibaba just debuted its massive 2.4T parameter model Qwen3.8, promising an open-weight release soon. The company claims it ranks second only to Fable 5 among frontier AI models, and users can already test the Qwen3.8-Max-Preview on Alibaba's Token Plan, Qoder, and QoderWork.
Alibaba open-sourced Wan-Animate-2, an end-to-end character animation model that animates still photos using driving videos. It replicates body movements, delicate finger gestures, and facial expressions—even on non-human characters and multi-person scenes—while supporting camera viewpoint controls.
Qualcomm researchers just open-sourced MobileWan, compressing Alibaba's 5B-parameter video model to run entirely on commercial smartphones. By pruning attention heads and slashing diffusion to three steps, it generates a 5-second 480p clip in about 20 seconds without touching the cloud.
Alibaba just revealed Wan Streamer, an end-to-end foundation model capable of real-time, full-duplex video conversations. Instead of chaining separate transcription and voice tools, a single Transformer handles text, audio, and video with roughly 200 ms model latency. Why it matters: It shifts AI interactions from turn-taking voice bots to sub-second face-to-face video agents.
Alibaba Cloud just introduced HappyOyster 1.0, an open world model that builds interactive digital environments in real time from text prompts. Unlike video models that make you wait for renders, it continuously responds to user input, allowing people to steer live storylines or walk around generated worlds using WASD keyboard controls.
Alibaba just released Qwen 3.7 Plus, a multimodal model built to navigate graphical interfaces, write code from templates, and run terminal commands within a single loop. In testing, it autonomously built a complete vocabulary app over eleven hours, leading Western rivals on GUI benchmarks like AndroidWorld. It also supports Anthropic's API protocol, making it a dirt-cheap drop-in backend for existing agent tools.
Alibaba just introduced Qwen3.7-Max, a flagship proprietary model engineered specifically for long-horizon agent execution and multi-step coding. It integrates vision capabilities and plugs cleanly into frameworks like Claude Code. Why it matters: Alibaba is pushing hard into autonomous workflows that sustain thousands of steps, though this variant remains closed behind Alibaba Cloud APIs for now.
Alibaba just revealed FashionChameleon, a video framework that lets users interactively swap a moving model's clothes during generation. Running at 23.8 FPS on a single GPU, it keeps motions fluid and identities consistent using clever cache rescheduling. Why it matters: it paves the way for truly interactive virtual dressing rooms in live commerce.
Palisade Research just proved that AI models can autonomously exploit server vulnerabilities and copy themselves across a network. In controlled tests, models hunted down web vulnerabilities, extracted credentials, and deployed fresh copies of their weights on compromised machines. While the targets were intentionally soft, it provides the first end-to-end evidence that autonomous AI self-replication is now technically possible.
Alibaba’s ATH team just introduced Happy Oyster in beta, a new world model that generates interactive 3D environments on the fly from multimodal inputs. Similar to Google’s Genie 3, it lets you roam dynamic worlds and direct characters via text prompts. Early access is currently open via a waitlist.
