Underdog releases Bark 3.0, a tiny open speech model that runs on phones
Underdog says Bark 3.0 is a 163 MB speech recognition model that runs on laptops and phones and is free and open under CC BY 4.0.
Positive AI on X
Ai2's FairStress tests whether AI stays fair under pressure
Ai2 introduced FairStress, an open benchmark that checks how readily a model changes a well supported decision based on demographics and how it explains itself.
Arc Institute's Bolero predicts how each cell reads the same DNA
Arc Institute introduced Bolero, a model that predicts chromatin accessibility and gene expression from DNA alone, conditioned on a cell's state.
DeepSeek open sources DeepInfer, the engine behind its V4.1 Flash service
DeepSeek released DeepInfer, the in-house inference framework it says powers DeepSeek-V4.1-Flash at scale, as open source to help lower AI compute costs.
Gemini 4 Argon builds a 3D periodic table that sings along to The Elements
Google DeepMind showed Gemini 4 Argon building an animated 3D periodic table synced to Tom Lehrer's song The Elements.
OpenAI acquires Orchestra Bio to help new treatments reach patients sooner
OpenAI acquired Orchestra Bio, a platform that helps biopharma teams coordinate drug development, to build on its Rosalind life sciences models.
Saint-Gobain is using Microsoft Discovery to develop more efficient building materials
Satya Nadella met Saint-Gobain in Paris and shared how the company uses Microsoft Discovery to help speed up development of more efficient construction materials.
ETH Zurich teaches a robotic hand to walk on its fingers
MIT CSAIL spotlighted ETH Zurich research in which a humanlike robotic hand learned to crawl, steer and recover from falls using its own fingers as legs.
Gemini team builds a playable 3D Wizard Chess game
The Gemini app team shared a playable 3D Wizard Chess game they built from scratch with Gemini 4 Argon and Google Antigravity in the Gemini app for Mac. The demo shows a multi-agent Blender workflow handling modeling, design and real-time animation.
Google.org backs 17 teams using AI for science with $36 million
Google Research highlighted the 17 nonprofits and academic teams sharing $36 million from the Google.org Impact Challenge: AI for Science.
Meta publishes a technical report on open model Muse Glimmer
Meta published a technical report on Muse Glimmer, a 30B open-weight model built for agentic tasks on consumer hardware. Meta says it runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision and leads similarly sized open models on tool use.
MiniMax launches M3.1 Flash, revealing it was Space Bunny
MiniMax launched M3.1 Flash and revealed that Space Bunny, the anonymous model that topped community leaderboards, was an early version of it. MiniMax says the model delivers frontier-level intelligence for agentic work at a lower cost.
OpenAI models get better at COBOL, Fortran and Visual Basic
OpenAI says its latest model advances are stronger at building and migrating software written in low-resource languages like COBOL, Fortran and Visual Basic. Those languages still run systems that banks, governments and businesses rely on, so help modernizing them could matter widely.
Viggle unveils JST-2, a world model built in 3D
Viggle says JST-2 builds explorable 3D worlds directly, generating a scene in seconds and expanding it as you move through it.
Anthropic commits shares to a new Public Benefit Fund
Anthropic announced the Anthropic Public Benefit Fund, a new wholly owned entity dedicated to the public benefit. It starts with 5% of the company's shares, with more set aside as the company's value grows past a set baseline.
DeepSeek ships V4.1 Flash 1017
DeepSeek released an updated V4.1 Flash model, now live on its website, app and API. The company says it posts stronger benchmark results, finishes tasks faster and gives shorter outputs than the previous version, which it describes as a gain in both performance and cost.
Claude designs leaner quantum circuits than the best published ones
Anthropic says Claude Mythos 5, working in a loop with an automated checker, found quantum circuit designs that beat the best published versions.
Four Gemini agents beat Minecraft on their own
Google DeepMind says a team of four Gemini 4 Argon agents defeated the Ender Dragon in about 27 hours, writing all their own code.
Google Research puts AI agents to work at the lab bench
Google Research connected AI agents to lab robots for fast design and test cycles, reporting tight binding proteins for a cancer related target.