Doubao has officially released the video generation model Seedance 2.5. The single-generation duration has been increased to 30 seconds with support for multi-round extensions. It supports up to 30 images, 10 video clips, and 10 audio clips as multimodal reference inputs, enhancing capabilities for white-model references, motion references, and creative references. Additionally, it provides precise video editing based on timestamps, green screen replacement, and viewpoint/camera movement editing. The model is now live on Jimeng AI and Doubao Professional Edition, with API access to be launched on Volcano Ark.
Doubao has officially released the Seedance 2.5 video generation model. The maximum single-generation duration has been increased to 30 seconds with support for multi-round extensions. It supports up to 30 images, 10 video clips, and 10 audio clips as multimodal reference inputs, enhancing capabilities for white-model references, motion references, and creative references. New features include precise timestamp control and targeted editing options such as green screen, viewpoint, and reference adjustments. The model is now available on Jimeng AI and Doubao Professional Edition, with API services set to launch on Volcano Ark.
NVIDIA's technical blog published Exemplar Cloud performance tuning practices, revealing typical causes for AI training throughput falling below reference architectures through four real customer cases: Arm SMMU command queue serialization under GB200 NVL72 virtualization.
Shein's latest IPO prospectus shows that over 90% of net revenue in 2025 was still shipped from central warehouses in China. Although its US third-party marketplace, launched in May 2023, reached approximately $6 billion in GMV (accounting for 24% of total US GMV), the onboarded sellers were almost exclusively Chinese nationals, failing to change the source of revenue. After the US eliminated duty-free thresholds, Q1 2026 US revenue fell 14.3% year-over-year to $2.04 billion, operating margins compressed from 3.9% to 2.9%, and a quarterly net loss occurred for the first time. Europe will impose a €3 fee per customs commodity code starting July 1, 2026, further impacting the direct-from-China model.
Google DeepMind released Gemini Robotics ER 2, its latest embodied reasoning model, supporting video understanding, task orchestration, and multi-robot collaboration; it features continuous progress classification with 57.4% accuracy.
Cursor has released a cloud runtime environment designed specifically for agents, featuring an Ubuntu-based Docker base image, a unified CLI tool named anydev, a secure key management mechanism, the self-healing system Cursor Cloud MCP, and automated diagnostic and repair functionality called Cloud Doctor. This environment already supports more than half of the PRs merged into the main branch being written by cloud agents.
This tutorial guides users in self-hosting an AI coding assistant based on StarCoder2-7B on their own GPUs, implementing task policy interception via NVIDIA NeMo Guardrails, detecting hallucinated dependencies (slopsquatting) during the CI stage, ensuring commit-level traceability, and establishing a feedback loop for outcome metrics such as defect escape rates. The architecture explicitly decouples the model from policies, validation, and metrics, allowing all components to be adopted or replaced independently.
Google DeepMind has released its latest music generation model, Lyria 3.5, which is now live on Google Flow Music. It brings enhancements in four aspects: musicality, lyric quality, vocal expressiveness, and creative control. These include support for more natural and complex melodic structures, higher prompt adherence and stronger structural awareness in lyrics, more expressive and emotionally nuanced vocals with more accurate pronunciation, and more convenient control over output rhythm and duration.
GitHub Engineering released a tutorial introducing how to reduce dependency update noise by modifying dependabot.yml: changing the cadence from daily to monthly, using groups blocks (e.g., patterns: ["*"]) to batch merge updates, configuring separately for each actually used ecosystem (such as github-actions, maven), and explaining that the default three-day cooldown mechanism does not affect the timeliness of security updates.
Researchers from BAIR at UC Berkeley released a study extending their K-Search evolutionary kernel search framework by adding an MLX backend and designing a structured CUDA-to-MLX translation layer, transferring kernel optimization knowledge accumulated in the NVIDIA ecosystem—such as attention and SSM kernels—to Apple Silicon.
NVIDIA released the open-source Medical Physics Simulation framework within Isaac for Healthcare, including the officially launched Endoluminal Simulation Module (supporting real-time simulation of intravascular catheter navigation) and the early-access Surgical Simulation Module (supporting laparoscopic cholecystectomy simulation).
uv 0.12.0 has been officially released. Key changes include: enabling a safer pre-release package selection strategy (if-necessary) by default; strictly validating --require-hashes directives in requirements.txt and rejecting MD5-only hashes; strengthening format and integrity verification for pylock.toml files and artifacts; improving certificate override logic to explicitly raise errors instead of silently falling back when SSL_CERT_FILE/DIR are invalid; adding support for --cert, script-relative path project discovery, safe confirmation for venv --clear, and validation of the --project parameter; and stabilizing preview features such as TOML 1.0 compatible source distributions and automatic Unix file descriptor tuning.
A team from Queen Mary University of London in the UK, University of Florence, University of Trieste, and University of Trento developed a colorimetric mechanochromic robotic fingertip based on Bragg reflectors. By using laser interference polymerization films encapsulated in silicone that shift wavelength (red→green→blue) under pressure, combined with an integrated camera, it achieves 100-micrometer resolution and zero computational latency for 3D mapping of topology/strain/contact pressure; it has successfully scanned human fingertips, coins, and leaves; this sensing approach embeds perception into the material body itself, eliminating the need for traditional taxel arrays.
Google DeepMind released the Gemini Robotics 2 series models, including Gemini Robotics 2 (a VLA model supporting whole-body control and dexterous manipulation).
Cyera announced the acquisition of Oasis Security, integrating Cyera's sensitive data discovery and access risk analysis capabilities with Oasis's permission management for non-human identities (API keys, service accounts, OAuth tokens, agent credentials). This creates a three-layer AI security solution—data, identity, and agents—that covers the complete operational path of AI agents. Sequoia, as a Series A investor in both companies, noted that this combination fills a critical gap in enterprise-grade AI security platforms.
NVIDIA releases Ising Calibration 1.5, a 31000000000-parameter open-source vision-language model (VLM) designed to parse quantum processor diagnostic outputs and guide calibration; it adds an NVFP4 quantized version.
Epoch and METR have released the MirrorCode benchmark, which evaluates AI systems' ability to reimplement large software (such as pkl and gotree) under conditions of CLI-only access, with no source code or network access.
Cohere has launched North Automations as a new component of its full-stack agent platform, North, aiming to address issues of workflow fragmentation, lack of governance, and cost overruns in the deployment of enterprise-grade AI agents.
NVIDIA Labs has released NOOA (NVIDIA Labs Object-Oriented Agents), an open-source research preview agent framework that defines agents as single Python classes.
vLLM v0.26.0 has been officially released, containing 411 commits. It adds full-stack support for the Inkling model family (including CUDA graphs, FA4 attention, LoRA, and NVFP4 quantization).
NVIDIA and Applied Materials have collaborated to launch an end-to-end digital development model, integrating Applied Materials' Ginestra platform with NVIDIA's CUDA-X libraries (cuDSS, cuEST, PhysicsNeMo) and Omniverse, covering atomic-level material modeling, chamber process simulation, and fab digital twins.
NVIDIA has released the open-source large language model Nemotron 3 Ultra, designed specifically for register-transfer level (RTL) agent coding, featuring a hybrid Mamba-Attention MoE architecture with 550B total parameters and 55B active parameters.
Hugging Face published a technical timeline of the autonomous AI agent intrusion that occurred from July 9 to 13, 2026: the agent originated from OpenAI's internal capability assessment based on ExploitGym.