Newest first
Hugging Face published a write-up on scaling pretraining for Indic languages to 2.8 trillion tokens, along with an open data generation pipeline. Few further details are available beyond the title.
Apertus 1.5, the open model, has release notes published on Hugging Face.
Open sourceBriefing3 sources
Meta says Muse Spark Open is the most capable US open-weight model, with 816 billion parameters and a 1 million token context, and runs locally on two Apple M5 Ultra Mac Studios.
Hugging Face, The Verge +11 min read
Hugging Face added support for the FLUX.3 Image Klein model in its Diffusers library.
OpenGCUA is an ongoing project to build a fully open stack for generalist computer-use agents spanning visual and programmatic interfaces, noting that frontier systems like GPT-6 Astra's remain closed.
MAI-Gym is an open platform unifying GUI agent evaluation, data collection and online RL training across mobile, computer and web environments, supporting ten benchmarks.
Liquid AI's LFM2.5-24B-A2B model is positioned for fast, private on-device agents.
SolidCraft is an open-source parametric CAD tool written in Rust with MCP support, letting AI agents drive it.
An experimental pattern-first digital audio workstation that supports MCP was released as an open-source project.
Janas is a GitHub project for running LLMs larger than RAM by streaming mixture-of-experts weights from SSD.
A GitHub project uses AI agents to reverse engineer software, from observing app behavior down to analyzing native binaries.