Alibaba open-sources Qwen3.8-Flash at about one-ninth training cost
Alibaba calls the 125B-parameter model a technical preview of its next-generation Qwen4 architecture.
Modelos, agentes, aplicaciones: la inteligencia artificial puesta a trabajar en problemas reales.
Alibaba calls the 125B-parameter model a technical preview of its next-generation Qwen4 architecture.
The anonymous model drew more than 50 trillion tokens in five days before Zhipu AI revealed its identity.
Gemini 3.5 Transcribe is already powering Gboard’s Rambler feature on Pixel 11.
Spare Parts 3D combines AI geometry extraction with engineer review.
HydroGym used simpler flows to train a controller that cut friction on a simulated wing by 38%.
CrysVCD applies chemical rules before AI generates crystal structures, reducing costly stability screening.
The index covers all Apple Top 200 podcasts across 135 verticals, adding 20,000 episodes each day.
Ring will begin a worldwide rollout in September, with TAKE as the default for cloud video features.
Two Peter Mac-led studies used AI to assess immune cells in more than 5,600 breast cancer patients.
Victoria University researchers trained two AI models on one million Reddit health-related threads.
OpenRouter counted around 295 million requests to Ox Alpha by August 25.
Granite 4.2 comes in 3B, 8B and 30B parameter versions for download and self-hosting.
Qwen 3.8 27B and PPLX 27B initially run on Nvidia's DGX Spark with 128 GB of Unified Memory.
NWPU’s FIRMED dataset marks emotional moments in four-second windows, rather than labeling an entire video.
It automatically organizes clinical conversations into SOAP-format electronic medical records.
The Indian startup processes 20 million call attempts a month and serves 1,200 Practo clinics.
Anthropic’s update lets Claude carry project details from conversations into Cowork without a fresh briefing.
Nvidia is distributing the new Ray Reconstruction model as “Preset F” through the Nvidia app.