Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
The Tongyi Qianwen team has officially released Qwen3.7-Max, a new-generation flagship large language model designed for the agent era, now available via API service through Alibaba Cloud Bailian.
Cohere has released the open-source large language model Command A+, which adopts a sparse MoE architecture with 218B total parameters and 25B active parameters, supporting a 128K input context window and multimodal capabilities (text, images).
AllenAI has released OlmoEarth v1.1, a more efficient family of Earth observation models that reduces computational costs by up to 3x while maintaining the performance of v1. The core improvement lies in modifying the token representation of Sentinel-2 data to reduce sequence length. Model weights, training code, and technical reports have been open-sourced.
Google Research published a paper in Nature introducing ERA, an empirical research assistant system developed by the company. The system uses Gemini to iteratively write and optimize scientific code, and has been applied in genomics.
OpenAI is advancing AI content provenance by launching Content Credentials, SynthID, and verification tools to help users identify and trust AI-generated media content.
The Tongyi Qianwen team released Qwen3.5-LiveTranslate-Flash, a multimodal simultaneous interpretation model built on Qwen3.5-Omni. It supports audio and visual input, covering 60 input languages and 29 output speech languages; end-to-end average character latency has been reduced to 2.8 seconds; it supports real-time voice cloning and hotword enhancement; and provides DashScope API access along with complete Python client examples.
Hugging Face's official blog announced the release of six new models in the Ettin Reranker series: from cross-encoder/ettin-reranker-17m-v1 to cross-encoder/ettin-reranker-1b-v1.
Google DeepMind has launched the Co-Scientist tool to help biologists accelerate research into reversing cellular aging. First, it identified over 20 novel, verifiable candidate genetic factors from literature; second, it compressed data-literature correlation analysis, which originally took six months, down to just a few days. Laboratories have verified that some of these factors can drive cells into a younger state with improved functionality.
PaddleOCR 3.5 adds native inference backend support for Hugging Face Transformers, allowing developers to call PP-OCRv5 via the engine='transformers' parameter.
Google DeepMind has integrated Google Street View into its experimental Project Genie prototype, enabling the generation of immersive virtual worlds based on real locations in the United States. The feature offers various styles (such as Ocean World and B&W film) and character customization capabilities; it is being gradually rolled out to global Google AI Ultra $200 subscribers alongside Project Genie.
Google DeepMind released Gemini Omni Flash, the first model in the Gemini Omni series. It supports high-quality video generation starting from video inputs, fusing image/audio/text inputs, and allows conversational editing via natural language. It has launched on the Gemini app, Google Flow, and YouTube Shorts. Future updates will support image and audio output. All generated videos embed SynthID digital watermarks.
Google DeepMind released Gemini for Science, which includes Hypothesis Generation (based on Co-Scientist).
Google DeepMind announced the expansion of its SynthID watermarking technology and C2PA Content Credentials to Search and Gemini.
SGLang v0.5.12 has been officially released, with core new features including complete inference path support for DeepSeek-V4, covering Tensor/Expert/Context parallelism and HiSparse KV offloading.
Google DeepMind has launched the Co-Scientist tool. Professor Clare Bryant of the University of Cambridge used it to analyze molecular switches that cause severe disease across different pathogens; the tool generates and ranks hypotheses based on research proposals, helping teams rapidly narrow experimental targets from candidate proteins to specific amino acids, with an expected reduction of the original 2–3 year experimental cycle to 6 months.
Google DeepMind launched the Co-Scientist tool to help researchers at the University of Edinburgh organize liver disease literature and generate new hypotheses. In studies on metabolic dysfunction-associated steatohepatitis (MASH), the tool proposed a hypothesis linking inflammation and metabolism via the NLRP3 inflammasome, which was subsequently validated by experiments.
A study by geneticist Gary Peltz's team at Stanford University, published in Advanced Science, used Google DeepMind's Co-Scientist to screen three candidate anti-liver fibrosis drugs from existing pharmaceutical literature. Experimental validation showed that two of them effectively blocked fibrosis and promoted hepatocyte regeneration, including the anticancer drug vorinostat (which inhibited 91% of scarring injury responses), while two drugs independently proposed by the researchers were ineffective.
WeatherNext, an AI weather model jointly developed by Google DeepMind and Google Research, integrates decades of global weather data with specialized data on extreme tropical cyclones to achieve the first synchronized high-precision forecast of both tropical cyclone tracks and intensity (including rapid intensification).
Google DeepMind has released the Gemini 3.5 Flash model, focusing on agentic and coding capabilities, with scores of 76.2% on Terminal-Bench 2.1 and 1656 Elo on GDPval-AA.
vLLM has released version v0.21.0, officially deprecating transformers v4 and requiring a C++20 compiler; it adds integration of KV offloading with the Hybrid Memory Allocator (HMA).
IBM Granite has released Granite Embedding Multilingual R2, which includes two open-source multilingual embedding models under the Apache 2.0 license: a 97M-parameter compact model (with an MTEB multilingual retrieval score of 60.3) and a 311M-parameter full-size model (with a score of 65.2).
A Hugging Face blog post details how asynchronous continuous batching is achieved by separating CPU batch preparation from GPU inference computation using CUDA streams and events. Empirical tests on an 8B model with a batch size of 32 generating 8K tokens showed that GPU utilization increased from 76.0% to 99.4%, while total latency decreased by 22% to 234.5 seconds. The complete implementation has been merged into the transformers library.
In a blog post, AI21 proposed a cache key design method for agentic LLM pipelines. By recursively encoding step names, local run indices, and upstream dependency keys into the cache_key, this approach supports result reproducibility while preserving necessary randomness within the same experiment. The article outlines three construction rules, example key formats, and Python implementation snippets, explaining their practical utility in A/B testing, best-of-N inference, and cost savings.
PyTorch 2.12.0 has been officially released. Key updates include up to a 100x performance improvement for batched linalg.eigh on CUDA, and the introduction of torch.accelerator.Graph for unified graph capture and replay.