News

Open-sourcing Rebalancer for solving resource allocation problems

Meta has open-sourced Rebalancer, a high-performance library used internally for over nine years to solve complex resource allocation and assignment problems. The framework separates problem specification from its solution to improve usability and scalability across hyperscale datacenters.

Base labs launches open-weight safety partnership

Baseten has introduced a new safety infrastructure standard through its Base Labs research arm. The initiative teams up with Hugging Face and Goodfire AI to monitor and secure open-weight models.

Also in Open Source AI

Suno releases v6 music model trained with record industry partnerships

Suno has launched its v6 music generation model, developed using licensed data from major record labels. The release introduces three model variants alongside new editing features and multi-modal prompting.

Also in Open Source AI

AlphaGenome Atlas maps every potential human DNA variant

A new platform provides precomputed predictions for nine billion single-letter DNA changes across the human genome. The resource aims to help researchers interpret genetic variations and accelerate discoveries in rare disease and population genetics.

Grok build harness released as open source

The development harness for Grok Build is now publicly available on GitHub. This release provides a complete reference for its agent loop, terminal interface, and extension system.

Nvidia releases open-source software to link idle home computers for AI processing

Nvidia has introduced PAIR, an open-source tool that connects idle household computers to handle complex artificial intelligence workflows. The software pools underutilized processing power across different devices without interrupting regular tasks.

Also in Open Source AI

Introducing Gemini 3.8 Flash and Flash Cyber

Google expands its family of fast AI models with Gemini 3.8 Flash, focusing on software engineering, complex reasoning, and multi-step agent tasks. A specialized cybersecurity variant, Flash Cyber, is also introduced for defensive security operations.

Also in Open Source AI

Introducing Grok Bot for autonomous task execution

Grok Bot provides autonomous AI teammates with their own virtual computers to complete multi-step work across applications. The system is available now in beta for select subscriber tiers on desktop and mobile devices.

Also in Open Source AI

TimesFM-3 model for multivariate time series forecasting

TimesFM-3 is a new time series foundation model designed to natively handle multivariate data. It processes complex relationships across multiple variables in a single pass.

Z.ai confirmed as the creator of Ox Alpha

Following weeks of speculation across developer communities, artificial intelligence laboratory Z.ai has confirmed it created the anonymous model Ox Alpha. The high-performing reasoning system is scheduled for a full weight release.

Also in Open Source AI

Metaroce: an open network protocol for ai-scale ethernet

Training massive artificial intelligence models requires moving data between thousands of graphics cards without wasting time or computing power. To solve this at scale, a new network protocol called MetaRoCE is designed to run efficiently on standard Ethernet connections.

Ornith Releases Version 1.5

Ornith has released version 1.5 of its open-weight models, featuring a self-improvement loop across three scales: a 397B mixture-of-experts flagship, a 35B mixture-of-experts model, and a 9B dense model.

Alibaba releases Qwen Video Edit tool

Alibaba has released Qwen Video Edit, a new AI video editing tool accompanied by a dedicated project page. The software provides video editing capabilities through instruction-based manipulation.

ByteDance releases Bernini v2 on Hugging Face

ByteDance released Bernini v2 on Hugging Face according to www.orcarouter.ai. ByteDance released Bernini v2 Diffusers on Hugging Face according to www.orcarouter.ai.

Also in Open Source AI

Tencent announces WorldClaw

Tencent Hunyuan has released WorldClaw, an AI-driven text-to-3D research paper and system designed to construct large, explorable virtual scenes. By utilizing planning agents alongside render-based refinement, the method aims to generate instance-level editable assets rather than a single locked render.

Sand AI launches Magi 2 preview model

Sand AI has reportedly launched the Magi 2 preview model on Hugging Face, introducing a 114 billion parameter unified audio-video generation architecture. Independent reporting from free.theresanaiforthat.com details the open-weights release, which activates approximately 6 billion parameters per token.