5 stories that day · 1 model · 3 open source · 1 research
Published daily at 7 PM ET · Updated · RSS
Open sourceTop story
Meta says no US company has shared a more capable open-weight model. It has 816 billion parameters, uses 42 billion per token and reads up to 1 million tokens; two M5 Ultra Mac Studios can run it.
Open source
Super VL has 120 billion parameters and understands images too; NVIDIA says it is five times faster than its closest rival in its class. Bolt 74B-A9B fits in 24 GB to run agents on your own machine.
NVIDIA · Original source (opens in a new tab)
Model
Krea releases Krea 3, an image model built for editing (opens the original source in a new tab)Krea says the model changes only the parts of a picture you ask for, and does best at swapping product scenes, removing objects and cleaning up photos. It is in the Krea app and API today.
Research
Microsoft AI finishes training MAI-Base-2, a 2.8-trillion-parameter model (opens the original source in a new tab)It is a base model, the raw version before tuning for chat, using 80 billion parameters per token. Microsoft says it beat every open-weight model it tested and is 4-12x more efficient to train than DeepSeek-V4-Pro-Base.
Microsoft AI · Original source (opens in a new tab)
Open source
Timestral predicts what comes next in a time series, such as sensor readings logged over time. Mistral says its largest, with 209 million parameters, ranks first on the field's three main public tests.
Mistral AI · Original source (opens in a new tab)
Notable stories from the past few days. Each shows its own date.
