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.
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'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.
Researchers just introduced SEGA, a training-free method that lets diffusion transformers generate coherent images far beyond their original training resolution. By dynamically scaling RoPE embeddings according to spatial-frequency energy at each denoising step, it balances global layout with fine textures. It means existing models like Flux can output 4096×4096 visuals without costly fine-tuning.
Alibaba just launched Qwen-Image-2.0-Pro, its latest text-to-image model featuring upgraded image quality, multilingual text rendering, and sharper instruction following. The new model debuted at #9 worldwide on Arena's Text-to-Image leaderboard, also cracking the top 10 for portraits and photorealistic styles. It is available to test on ModelScope alongside an official API.
Alibaba Qwen's central technical leader Junyang Lin just announced he is stepping down from the open-weight project, only a day after releasing the Qwen 3.5 small models. Colleagues called the abrupt exit an immense loss, with one teammate hinting the departure wasn't Lin's choice. The shakeup comes right as China's top open-source model family gains global momentum.
Alibaba's Qwen team just rolled out Qwen3.5-397B-A17B, the first open-weight model in the Qwen3.5 lineup. Built on a sparse MoE architecture, it carries 397 billion total parameters while activating just 17 billion per token, pushing decoding throughput up to 19x over Qwen3-Max. Why it matters: Developers get Apache 2.0-licensed multimodal agent capabilities without the prohibitive serving costs of dense giants.
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, just unveiled Tinker, an automated tool designed to fine-tune frontier open-source models. Developers can use a few lines of code to run supervised fine-tuning or reinforcement learning on Meta's Llama and Alibaba's Qwen, bypassing messy GPU cluster management entirely. Access is open via application for free during beta.
Qwen just introduced Qwen-Image-Edit, a 20B-parameter foundation model built specifically for advanced image and text editing. It can directly modify Chinese and English text inside images while preserving fonts, alongside pulling off full 180-degree novel view synthesis and localized element tweaks. You can already try it inside Qwen Chat.
Top Chinese AI apps like ByteDance’s Doubao, Alibaba’s Qwen, and DeepSeek just disabled photo-recognition and exam-solving features during the country's college entrance exams. Chatbots told users the temporary blackout was implemented to prevent cheating. Basically, an AI kill-switch for exam season.
BMW and Alibaba just expanded their partnership to bring Alibaba's Qwen LLM into BMW's Neue Klasse electric vehicles in China starting in 2026. The co-developed AI engine powers dual in-car agents capable of handling complex multi-intent navigation and lifestyle queries. Why it matters: Global automakers are increasingly relying on local AI giants to survive China's fiercely competitive smart EV market.
Manus AI just partnered with Alibaba's Qwen team to integrate its viral general AI agent with Qwen's open-source models and domestic computing infrastructure. After facing massive hype and severe site crashes following its launch last week, Manus is leaning on Alibaba's resources to manage traffic surges and scale beyond its current invite-only rollout.
Butterfly Effect's new autonomous agent Manus just went viral, with invite codes fetching thousands of dollars after claims it outperforms OpenAI's tools. But beneath the hype and multi-model routing (including Claude and Qwen), early testers find it frequently crashes, hits endless loops, and fails basic tasks like ordering food or booking flights. The team says it's still in closed beta.
Qwen just launched QwQ, a mid-sized reasoning model that matches heavyweight frontier systems like DeepSeek-R1. Alibaba found that scaling reinforcement learning yields massive leaps in math and coding, proving mid-sized models can punch way above their weight against massive MoE giants.
Qwen just unveiled QwQ-Max-Preview, a preview of its upcoming reasoning model built on Qwen2.5-Max that targets math, coding, and autonomous agents. The team also pledged to open-source the full QwQ-Max under Apache 2.0, along with a smaller QwQ-32B designed for local device deployment.
The Qwen team just released the full Qwen2.5-Coder lineup, spanning six sizes from 0.5B to 32B parameters. The flagship 32B-Instruct model matches GPT-4o on major coding benchmarks like EvalPlus and Aider, supporting 40+ languages and 128K context. Most models run on an Apache 2.0 license, bringing serious frontier-level coding power directly to local hardware and Cursor setups.
