What happened in AI today

More stories

1. NVIDIA and AI4Bharat announce 2.8 trillion token Indic pretraining dataset

   Hugging Face · Open source

   NVIDIA and AI4Bharat, the Indian language research lab at IIT Madras, announced Nemotron-pretraining-Indic-v1, a corpus of about 2.8 trillion tokens covering the 22 languages listed in India's Constitution. It has native-script data for all 22 languages and romanized text for twelve. The release is due soon, along with the open, GPU-accelerated pipeline that made it, which includes the prompts, LLM workflows and quality checks. The data is translated from curated English sources in Nemotron-CC-v2, and NVIDIA says it complements naturally written Indic text rather than replacing it. NVIDIA says 10 billion Indic tokens drawn from the corpus and natural text raised scores on five Indic benchmarks for Nemotron 3.5 Super VL by 4.98 to 9.82 percentage points. That result covers the whole training stage, not the synthetic data alone.

   • Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline

2. Swiss AI releases Apertus 1.5 open models with image and audio input

   Hugging Face · Open source

   Swiss AI has released Apertus 1.5, an update to its 8B and 70B Apertus 1.0 models. The new version accepts images and audio alongside text, has an optional thinking mode, and has a 262,144-token context window, four times larger than Apertus 1.0. The weights were released to Hugging Face on July 24, 2026 under an Apache 2.0 license, and the team says audio input is still experimental. Swiss AI says Proton Lumo and Infomaniak Euria have integrated the models, and it has announced a roadmap for Apertus 2.0 with a sparse/MoE design and about 600B total parameters.

   • Apertus 1.5 Release Notes

1. Liquid AI releases open d1 decision models for edge devices

   Hugging Face · Open source

   Liquid AI has released two open-weight decision models on Hugging Face, d1-3B and d1-omni-600M. Decision models answer in a single forward pass instead of producing tokens. Liquid AI says d1-3B scores 48.57 on the Decision Index 0.2.1, ahead of every 4B and 9B model and of Decider 35B-A3B (47.11), and answers one question in 16 ms on an NVIDIA Jetson AGX Thor. The d1-3B model takes text and images, while d1-omni-600M, an experimental early research release, takes text with either images or audio. Liquid AI reports a mean score of 82.9 for d1-3B and 78.4 for d1-omni-600M across seven public datasets.

   • Multimodal open d1 decision models for the edge

1. Cohere releases Tiny Aya L2-Thinker, a small model that reasons in the user's language

   Cohere · Open source

   Cohere released Tiny Aya L2-Thinker, a 3.35B-parameter model, along with its multilingual reasoning data on Hugging Face. Cohere says the model reasons in the prompt language over 93% of the time across 60 languages, with minimal accuracy loss and fewer reasoning tokens than comparable models. Most reasoning models think in English, which Cohere says can lose nuance and leave non-English speakers unable to read the reasoning. Cohere says its method mixes broad-language reasoning data, a small share of multilingual non-reasoning data, and English reasoning data. The release also includes Tiny Aya En-Thinker and Tiny Aya Base 32K.

   • Multilingual Bridges: How Data Mixing Unlocks In-Language Reasoning

1. Ai2 open-sources AstaBrief, an 8B model for cited research reports

   Ai2 · Open source

   Ai2 has released AstaBrief 8B, a model that turns a research question and retrieved literature excerpts into a cited report, along with its training data. The model is built on Qwen3-8B and is trained with supervised fine-tuning and direct preference optimization. Ai2 says it writes the full report in one pass and averages 51.1 seconds per report in Asta's Fast mode, against 178.5 seconds for the Claude-powered Thinking mode. On Ai2's SQABench-CS2 test set, AstaBrief scored an average of 88.8 against 86.2 for Thinking mode, and Ai2 says most training and evaluation work was done in 2025. AstaBrief is available in Asta's Generate a report feature as Fast mode and for download on Hugging Face.

   • Open-sourcing AstaBrief, the fast report-generation model in Asta

1. Ai2 releases Olmo-core 3, open training code for trillion-parameter MoE models

   Hugging Face · Open source

   Ai2 has released Olmo-core 3, an upgrade to its framework for building large language models with a redesigned mixture-of-experts (MoE) training system. Ai2 says it is designed to scale MoE training into the trillion-parameter range. In one benchmark, Ai2 grew the expert pool from 8 to 128 and total parameters from 4.6B to 47B, while training throughput fell by less than 5%. In a preliminary test on eight NVIDIA B300 GPUs, Ai2 reports 2.7 times the throughput of its earlier implementation, and says it benchmarked a 1.2-trillion-parameter model across 512 GPUs. The code is open on GitHub, and Ai2 says its next-generation Olmo will use an MoE architecture.

   • Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs
   • Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs

1. Qwen open-sources Qwen-Image-2.1 for image generation and editing

   Alibaba · Open source

   Qwen has open-sourced Qwen-Image-2.1, a unified text-to-image generation and image editing model. The visual generation component has 7B parameters and is designed to balance image quality with inference efficiency. Qwen says it can generate regular and transparent RGBA images, edit images using up to 10 reference images, and handle annotated local edits while improving text rendering and portrait detail.

   • Qwen/Qwen-Image-2.1

1. Nvidia releases open Kumo Tabular AI for spreadsheet-style predictions without training

   Hugging Face · Open source

   Nvidia says Kumo Tabular is an open foundation model that predicts labels for new table rows in a single forward pass with no training, tuning or feature engineering. It comes in three sizes from 28M to 215M parameters, was pretrained only on artificial data, and is available on Hugging Face with code on GitHub under the OpenMDW-1.1 license for commercial use. Nvidia says it ranks first on TabArena, BeyondArena, TALENT and ScoringBench while running faster than rivals.

   • NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction

1. Hcompany releases Holo4 agent models that use screens, code and APIs

   Hugging Face · Open source

   Hcompany introduced Holo4, a series of agentic models in 27B dense and 35B-A3B mixture-of-experts sizes for computer use. Hcompany says the same model can click and type in GUIs, write and run code, and call MCP tools or APIs across desktop, web, Android and business software. Both sizes are available on the H Models API with weights posted on Hugging Face, and Hcompany reports 61.7% on OSWorld 2.0 for the 27B model at much lower cost than frontier closed models.

   • Holo4: powering generalist computer-use agents

2. GitHub says its open-source AI security agent helped find 24 Android app flaws

   GitHub · Open source

   GitHub says researchers used its Security Lab Taskflow Agent with targeted Android audit prompts to find and report 24 vulnerabilities in Android apps. The post describes an OsmAnd navigation flaw that GitHub says let any app silently change settings to leak location and routes, and a Wikipedia app deep-link and cookie flaw that enabled account takeover. The taskflows are open source and can be run via Codespaces with a Copilot license, though GitHub says human review is still needed because the AI can misjudge severity and produce false positives.

   • How we found 24 Android vulnerabilities using our open source AI security agent

1. Nvidia releases Isaac ROS 5.0 to let AI agents help build robots

   NVIDIA · Open source

   Nvidia says Isaac ROS 5.0, released at ROSCon in Toronto, brings AI-agent workflows to the nearly 1.3 million-user ROS ecosystem with GPU-accelerated, open-source packages. It adds support for ROS Lyrical and Ubuntu 24.04, reusable skills for setup and manipulation, a FoundationStereo fine-tuning skill, and an agent-ready FoundationPose library Nvidia says tracks objects up to 5.5 times faster. The release runs from Jetson Orin Nano to Jetson Thor and is free and open source.

   • NVIDIA Isaac ROS 5.0 Advances Agentic, Open Source Robotics Development

2. Black Forest Labs releases open 7B robot-action model topping RoboLab

   Black Forest Labs · Open source

   Black Forest Labs says FLUX 3 Action is an open-weight 7B world-action model derived from its FLUX 3 backbone and pretrained largely on video plus image and audio data. It reports 38.3 percent success on RoboLab-120 for its single-step checkpoint and 42.2 percent for its guidance-distilled version, ahead of 36.8 percent for Cosmos 3 Nano, at less than half the parameters. The company also says it runs up to 3.95 times faster and has published the weights.

   • Introducing FLUX 3 Action