Doubao has officially launched the native audio-video full-duplex large model SeedRealtime. It employs a unified architecture that natively integrates audio, video, and text, achieving real-time interaction of seeing, listening, and speaking simultaneously on continuous multimodal streams. The model features three breakthroughs: joint audio-video understanding, proactive interaction capabilities, and smooth dialogue rhythm. End-to-end human evaluations show that dialogue rhythm issues are reduced by half compared to cascaded models. Currently fully rolled out in the Doubao App, users can experience it via the "Call" option in the video call interface after updating to the latest version.
Doubao officially launched the native audio-video full-duplex large model SeedRealtime. It adopts a unified end-to-end architecture that integrates audio, video, and text, achieving real-time perception, understanding, and expression on continuous multimodal streams. The feature is fully available in the Doubao App; users can experience it by selecting "Call" in the video call interface after updating to the latest version.
NVIDIA released the open-source 34 billion parameter Alpamayo 2 Super model, which integrates a 32 billion parameter Cosmos 3 Super Reasoner with a 2 billion parameter diffusion-based Action Expert.
Mistral has released Shieldstral, a 3B-parameter, Apache 2.0 open-source multimodal safety classifier that models content moderation as a policy-adaptive question-answering task; it supports querying in natural language at inference time (e.g., 'Is this image suitable for minors?'), handles text and images uniformly without retraining, and can run on a single 16GB NVIDIA GPU.
On July 28, the US Federal Communications Commission (FCC) added mobile robots weighing over 2 kilograms with communication capabilities to its 'Covered List', prohibiting the import of newly produced foreign devices of this type; exemptions include drones, connected vehicles, and medical devices.
A study by Diyi Yang's team at Stanford University, published in Nature Human Behavior, analyzed surveys from 1,131 Character.AI users and 244 complete chat logs. It found that users with smaller social networks, those primarily motivated by companionship, and those more willing to disclose sensitive information to AI reported lower subjective well-being. The research likens AI companionship to 'social snacking,' noting its lack of reciprocity and depth compared to real human interaction, which may create a vicious cycle of worsening loneliness.
The Stanford Institute for Human-Centered Artificial Intelligence (HAI) released a policy brief that systematically proposed the concept of "world models" for the first time—defined as AI systems capable of constructing internal representations of physical environments and predicting the consequences of actions—and categorized them into three types: renderers, simulators, and planners.
Google Open Source released its 2025 annual open source review, noting that approximately 10% of Alphabet's full-time employees are actively participating in open source projects, with contributions spanning communities such as LLVM, vLLM, and Kubernetes, as well as internal projects. It promoted the Agent2Agent (A2A) protocol and donated it to the Linux Foundation; released the Gemma 3 lightweight model and the Agent Development Kit open source framework; launched security tools including CodeMender, OSS Rebuild, OSV-Scalibr, and deps.dev credential scanning; invested $2 million across more than 40 projects through its OSPO, and supported 1,280 newcomers via Google Summer of Code.
Microsoft Research has released the open-source framework Orchard, whose core is Orchard Env—a lightweight, Kubernetes-native, reusable environment service that supports software engineering (Orchard-SWE).
An NVIDIA technical blog post published a tutorial introducing how to build three fully isolated tenant-level Kubernetes clusters on a single host equipped with an NVIDIA L40S GPU using KAI Scheduler v0.16.4 and vCluster CLI 0.35.1.
NVIDIA released the BlueField-4 STX storage processor in a technical blog post. Its built-in Vera CPU (based on an 88-core Olympus architecture) accelerates AES-128 encryption/decryption and Reed-Solomon recovery.
AMD officially launched the CDNA™ 5 architecture, Instinct™ MI455X GPU, and Helios™ rack-scale AI solution, integrating 72 GPUs into a single rack using the OCP Open Rack Wide standard.
Researchers from the University of Toronto, Vector Institute, Cambridge University, and ServiceNow have built a self-sustaining AI worm prototype based on open-weight LLMs, which uses compromised GPU nodes to run models for vulnerability detection and customized attacks.
MiniMax officially announced the open-sourcing of H3, its next-generation general-purpose video generation model. It supports unified understanding and generation of text, images, videos, and audio, capable of outputting videos with up to 2K resolution, 15-second duration, and native stereo sound.
The Tongyi Qianwen team officially released Qwen3.8-Max, with a total parameter count of 2.4 trillion (95B active parameters). It is the most powerful model in the Qwen family to date and the first Max-level model with open-sourced weights; it has been launched on the Qianwen AI platform and supports API calls.
The GitHub engineering team has open-sourced the Rust crate casefold for high-performance source code case folding. Its core optimizations include replacing early exits with branch-free full-buffer scans, achieving >45 GiB/s on ASCII paths (close to memory bandwidth).
NVIDIA released Video Codec SDK 13.1, adding AV1 hierarchical reference mode (supporting up to 31 B-frames), UHQ tuning information with iterative encoding fusion, per-macroblock decoding statistics, frame-level precise seeking API, MV-HEVC stereo enhancement, modular queue-based transcoding architecture, CUarray zero-copy transcoding, and an official Docker-based development environment.
OpenAI cracked down on a scam ring based in Cambodia that used ChatGPT to support criminal activities including investment scams, romance scams, gambling scams, and identity impersonation.
MiniMax has released the H3 multimodal generation model, which supports unified context understanding across text, images, videos, and audio. It can generate videos up to 15 seconds long at 2K resolution with native stereo sound, featuring instruction following and precise image-text brand rendering.
Based on six months of red team assessments across various types of AI agents, NVIDIA's AI Red Team identified four common security failure modes: lack of agent access control, tools supporting arbitrary code execution, absence of network egress controls, and plaintext exposure of keys. The article analyzes each risk category individually and provides deterministic architectural control recommendations, including mandatory access control, sandboxed command execution, default-deny network egress, and isolation of keys from the agent runtime environment. It emphasizes that internal model control mechanisms such as prompts or LLM-as-a-judge generally fail under adversarial pressure.
Google Research published a paper proposing the Chain-of-Evidence (CoE) framework, which defines structured standards for the credibility of research outputs: every claim must be accompanied by a complete and correct chain of evidence.
ByteDance officially released the video generation model Seedance 2.5, increasing the single generation duration to 30 seconds with support for multi-round extension, optimizing shot transitions and narrative coherence; it supports multimodal reference inputs of up to 30 images, 10 video clips, and 10 audio clips; new precise timestamp control and various editing capabilities including green screen, perspective, and reference are added; it is now live on Jimeng AI and Doubao Professional Edition, with API services to be launched on Volcano Ark.