Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
ElevenLabs has launched Image & Video (Beta), integrating video models such as Veo, Sora, Kling, Wan, and Seedance, as well as image models including Nanobanana, Flux Kontext, GPT Image, and Seedream. It also supports lip-syncing, Studio audio synthesis, music and sound effect layering, timeline editing, and export.
SGLang Model Gateway released version v0.2.3, introducing bucket-mode routing that improves TTFT and throughput by 20–30%, adding PostgreSQL chat history storage support, and enhancing support for the MinMax M2 model, structured outputs, Tool Choice streaming parsing, and OutputItemDone events.
SGLang 发布 Model Gateway v0.2.2，新增行业首个支持 Llama、DeepSeek、Qwen 等全部开源模型的 Responses API，集成 MCP、原生 gRPC 健康检查、mTLS 加密通信、Harmony Pipeline 统一架构，并全面支持 Linux/Windows/Mac/x86/ARM 及 Python 3.8–3.14。
SGLang officially released Model Gateway version 0.2.1, focusing on stability and performance. New features include a two-level tokenizer cache (L0 exact match + L1 prefix-aware), an OpenAI-style /v1/classifications API, and a Worker management workflow engine. Additionally, it fixes multiple issues such as UTF-8 boundaries, gRPC timeouts, and worker filtering, while completing documentation updates and code cleanup.
SGLang released Model Gateway v0.2.0 (formerly SGLang Router), introducing the Inference Gateway (IGW) mode for multi-model inference, which supports dynamic model registration, label-based routing policies, and health checks.
Hugging Face has launched the kernels library, which supports convenient building, testing, and publishing of high-performance kernels for AMD GPUs (ROCm) through the kernel-builder tool and the kernels-community collaborative ecosystem.
Google Research and Google DeepMind jointly released the Natural Forests of the World 2020 map and dataset.
Google DeepMind has released SIMA 2, which uses the Gemini model as its core, upgrading the agent from instruction-following to possessing capabilities in goal reasoning, multilingual/emoji understanding, cross-game generalization (e.g., ASKA, MineDojo), and iterative improvement based on self-generated experiences. It can autonomously navigate and execute tasks in entirely new 3D worlds generated by Genie 3, operating via keyboard and mouse interfaces within a limited context window. Currently, it is available as a limited research preview for academics and game developers.
GPT-5.1 is now available via API, offering faster adaptive reasoning, extended prompt caching, improved coding performance, and new apply_patch and shell tools.
ChatGPT has launched a group chat feature, enabling users to collaborate with others and ChatGPT within the same conversation.
Tongyi Qwen officially released DeepResearch version 2511, introducing Normal/Advanced dual modes, fused parsing of local PDF/Excel/image files, enhanced search reading, precise and controllable reports, a new interactive experience, and diverse output formats including web pages and podcasts. It employs a multi-agent collaboration mechanism, achieving comprehensive improvements over version 2507 in dimensions such as hallucination rate, report comprehensiveness, credibility, instruction following, report length, search depth, and research time.
Google Research has released JAX-Privacy 1.0, a differential privacy (DP) machine learning toolkit built on JAX. It supports algorithms such as DP-SGD, DP-FTRL, and DP matrix factorization, integrates micro-batching, cross-accelerator parallelism, and tight privacy loss accounting, and provides complete fine-tuning examples for the Gemma model series. The code is open-sourced on GitHub, with a PIP package also released.
OpenAI is opposing the New York Times' request to obtain 20,000,000 private ChatGPT conversation records and is accelerating the implementation of new security and privacy protection measures.
ElevenLabs has released the Scribe v2 Realtime model for real-time speech-to-text conversion. It supports 90 languages, including English, French, German, Italian, Spanish, and Portuguese. The model features end-to-end latency below 150ms and an average accuracy of 93.5% across 30 Eurasian languages. Key capabilities include negative latency prediction, automatic language detection, text conditioning, Voice Activity Detection (VAD), and manual submission. It is now available via API and integrated into ElevenLabs Agents.
DeepMind published a new paper in Nature proposing a three-step visual model alignment method: first, train a small adapter using the THINGS dataset to build a teacher model; second, generate a large-scale AligNet dataset; finally, fine-tune the student model. Experiments show that aligned models significantly improve performance on cognitive tasks such as odd-one-out and multi-image arrangement, as well as AI tasks like few-shot learning and distribution shift. Their internal representation structures are also clearly organized according to human conceptual hierarchies (e.g., animals, food).
SGLang v0.5.5 has been officially released. Key updates include Day 0 support for Moonshot AI's Kimi-K2-Thinking and MiniMax's MiniMax-M2 models, the addition of video and image generation capabilities, kernel optimizations for the Blackwell architecture, MoE runtime refactoring, and the introduction of overlap support for spec and prefill CUDA graphs.
Google Research has released DS-STAR, a new type of data science agent featuring three innovations: a file analysis module that automatically parses heterogeneous data formats (such as JSON, text, and Markdown); an LLM-based step-by-step plan verification mechanism; and a sequential planning process that iteratively optimizes based on feedback.
Moonshot AI has released Kimi K2 Thinking, claiming it is the current best open-source reasoning model. It supports up to 200–300 consecutive tool calls, has launched a chat mode at kimi.ai, with full agentic mode coming soon, and offers API access.
Moonshot AI has released Kimi K2 Thinking, claiming it is the current best open-source reasoning model. It supports up to 200–300 consecutive tool calls, is now available in chat mode on kimi.com, with a full agentic mode coming soon, and offers API access.
On November 5, 2025, Google DeepMind released three biosphere AI studies: first, a deforestation risk prediction model based on pure satellite inputs and Vision Transformer architecture, supporting large-scale predictions at 30-meter resolution.
Google Research and Google DeepMind jointly released the ForestCast model and the first public benchmark dataset for deforestation risk prediction, using raw Landsat and Sentinel-2 satellite data and 'change history' as the sole inputs.
The Berkeley Artificial Intelligence Research lab published a blog post introducing a new reinforcement learning algorithm, Transitive RL (TRL), which uses a 'divide-and-conquer' paradigm to replace traditional temporal difference (TD) learning, specifically for offline goal-conditioned RL.
At the Research@MTV event, Google Research announced three latest research breakthroughs: Earth AI integrates multi-source geospatial data with LLM reasoning capabilities, expanding its river flood model coverage from 700 million to over 2 billion people.
Google Research has proposed Verifiable Private Insights (PPI), a novel analytics framework integrating large language models (Gemma 3 4B), differential privacy (DP), and trusted execution environments (TEEs) to generate insights on GenAI tool usage while protecting users' raw data. The system has been deployed in the Pixel Recorder app, and all key components—including the confidential federated analytics (CFA) stack, the LLM structured summarization pipeline, and the TEE verification mechanism—have been open-sourced.