Google DeepMind has released Gemini Omni 1.1 Flash, adding features including scene extension (supporting up to 10 seconds of context, extendable cumulatively to 40 seconds), first-and-last frame interpolation, 360p fast preview (60% faster than 720p with two-thirds lower cost), 4K upscaling, and video reference input of up to 3 seconds. It is now available via the Gemini API on Google AI Studio and Agent Platform.
Cohere releases Parse, a vision-language model dedicated to large-scale enterprise document processing, supporting nine languages and outputting structured Markdown; priced at $1.50 per thousand pages via the Cohere API.
Google DeepMind, in collaboration with the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons, conducted the world's first double-blind AI evaluation of the Gemini Flash Lite model within a cryptographic environment, preventing models from accessing test questions prematurely and causing benchmark contamination.
Anthropic has launched a Claude team plan for scientists, offering 10,000 free standard seats and premium seats with 5x usage at $15 per month for one year. Simultaneously, it is expanding its AI for Science program by adding credit support for computationally intensive research fields such as mathematics and physics. The announcement also notes that biology and chemistry researchers are currently restricted from using Opus-class models, while Fable models continue to block queries related to professional biology and drug development.
Anthropic has released a research preview of the Model Hardware Standard (MHS), a shared specification for AI agents to safely operate physical devices, supporting parallel control of equipment such as microscopes, liquid handlers, and robotic arms. MHS enables cross-vendor device discovery and safe operation through standardized drivers, natural language labels, and protocols like MCP, and has been validated across multiple scenarios at institutions including Genentech, HHMI Janelia, and QuEra.
NVIDIA has released the NVLink Fusion connectivity technology and the accompanying NVHBM high-bandwidth memory substrate technology, enabling hyperscalers and AI-native companies to integrate their custom XPUs into NVIDIA's AI infrastructure platform; compared to standard HBM4e, NVHBM offers up to 30% higher memory bandwidth, up to 25% more compute die area, and up to 15% lower HBM power consumption.
NVIDIA's technical blog published a tutorial introducing how to use the AI agent-driven COMPASS (Cross-Embodiment Mobility Policy via Residual RL and Skill Synthesis) workflow.
Google Research introduced GlucoFM, a self-supervised foundation model for continuous glucose monitoring (CGM), which employs a dual-stream encoder to separate slow baseline trends from short-term deviations while preserving time-point and missing information.
Alibaba has released a preview version of the Qwen3.8-Flash-Next model, serving as a precursor to the upcoming Qwen4 architecture; this model features a multimodal MoE structure with a main model of 125B parameters plus 51B N-gram embeddings.
Google DeepMind has released Gemini 3.5 Transcribe, a new speech model designed for real-time intelligent speech transcription, supporting two API access modes: streaming (gemini-3.5-transcribe-live) and non-streaming (gemini-3.5-transcribe).
AMD has released ROCm 10.0, a major version update marking the tenth anniversary of its open-source GPU computing platform. Key changes include full reliance on the automated build system TheRock for releases and unified software distribution via repo.amd.com.
Zhipu AI has officially open-sourced GLM-5.3-Flash (320B total parameters, 18B active parameters), the first native multimodal model in the GLM-5 series, which scored 57 on the AA Comprehensive Intelligence Index, matching Claude Opus 4.8.
The Tongyi Qianwen team has released the open-source multimodal MoE model Qwen3.8-Flash-Next, with 125B parameters (including an additional 51B N-gram Embedding) and 6B active parameters per token.
Goodfire Labs released the Silico platform, providing AI model analysis tools based on mechanistic interpretability. The platform supports users in asking questions via natural language, with AI agents automatically designing experiments, executing tasks, locating neuron behaviors, and even performing direct edits. The platform is now open to the public, and a $1 million Silico usage credit grant program for academic and non-profit research has been launched simultaneously.
The vLLM project released version v0.28.0, focusing on optimizing end-to-end inference performance for the Kimi-K3 and DeepSeek V4 models, and adding support for AMD ROCm (gfx11/gfx950).
The Hugging Face official blog introduces the new MultiVectorEncoder model type in Sentence Transformers v6.0, which supports ColBERT-style late-interaction retrieval.
NVIDIA Dynamo has introduced the Shadow Engine Recovery preview feature, which enables second-level failure recovery for LLM inference services through the GPU Memory Service (GMS) and pre-initialized shadow engines.
Google Research unveiled the AgentHands research prototype at CHI 2026. By mapping large language model outputs to 3D gestures strictly synchronized with speech, it enables spatially anchored agent conversations in XR environments. The system comprises four key modules: environment perception, a gesture event library, embedded reasoning, and synchronous execution. Validation on orchid care and 3D printer operation tasks demonstrated significant improvements in spatial localization, action understanding, and safety prompts through gesture integration.
The IBM Granite team released the Granite 4.2 series of dense decoder-only reasoning large language models, available in 3B, 8B, and 30B sizes, post-trained on the Granite-4.1 base model.
Multiverse Computing and Hugging Face jointly released the paper "Quantization-Aware Healing," proposing the QAH method: For models that have undergone structural compression (e.g., GPT-OSS 120B→60B) and MXFP4 quantization, distillation is performed directly from the original full-precision, full-size teacher model, rather than from a restored bfloat16 checkpoint.
OpenAI banned accounts originating from Russia that used AI to promote a fictitious Israeli think tank and a "sovereignty" index that praised Russia and criticized the West.
Gradio has officially released its built-in gr.Workflow feature, which supports orchestrating typed node graphical AI workflows via a drag-and-drop canvas. Each node can be run independently with real-time visibility of intermediate results; the same workflow automatically provides a REST API and supports one-click deployment to Hugging Face Spaces. Examples include image editing, AI media studios, parallel image generation, dataset analysis, and local execution of GPU models.