Tencent's new open-weight text model has 770 billion total parameters (49 billion active at once) and a 1 million token context window, roughly doubling the size of its predecessor from just last month.

The updated video generation model can now extend scenes using 10 seconds of prior context, lets creators pin first and last frames to control camera movement, and supports 4K upscaling.

The new model ships as two endpoints — a fast streaming version and a more accurate batch version — reporting a 2.6% average word error rate across more than 85 languages.

Pollen Robotics' small bipedal robot uses neural network movement policies trained in simulation and exported for real-world use, with an Apache-2.0 licensed training stack anyone can retrain.

The new standard gives AI agents a unified way to interface with robotic arms and lab instruments, cutting integration time from weeks to hours in early tests, though human oversight remains necessary since models still struggle with physical cause and effect.

Z.ai and Qwen separately built nearly identical designs for their new 'Flash' models, both using a 3:1 hybrid attention mix, compressed indexing, and the Muon training method.

The Big Video Dataset contains 80 million videos with auto-generated descriptions, and models trained on it have already beaten the previous leading benchmark by up to 2.1 percentage points.

The framework lets AI agents document what worked and what didn't in a wiki-like knowledge base, and researchers found smaller models using it can match larger models that lack this memory.

The open-source Linux distribution's community held a general vote on how contributors may use large language models in the project, sparking a large debate.

The author argues that leaning on AI coding assistants has made software development feel less engaging and is dulling the deep technical know-how ('savviness') developers used to build through hands-on practice.

The piece questions whether large language models are genuinely getting smarter or simply becoming better at appearing capable, amid broader concern about an AI market bubble.

The project lets musicians and hobbyists isolate vocals, drums, bass, piano, and guitar from tracks locally for practice, remixing, or transcription, and has been gaining traction on Hacker News.

The TypeScript library treats shader files as importable modules that run identically in browsers, headless Node.js, and CI testing, and ships a full-screen visual effect in just 25KB.

Researchers who normally build tools to detect counterfeit medicines are now applying similar AI-based techniques to identify fake cosmetics.