Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
OpenAI has banned accounts that used AI to support romance scam workflows, including stages such as outreach, translation, victim interaction, and investment fraud inducement.
OpenAI has banned a batch of accounts associated with the suspected Russian-backed 'No Bell' operation, which uses AI to generate critical content targeting the United States and its allies for African audiences.
Google DeepMind has opened the Project Genie experimental prototype, based on the Genie 3 world model, to US Google AI Ultra subscribers (18+).
vLLM has released version v0.15.0, adding support for models such as Kimi-K2.5, Molmo2, Step3vl 10B, GLM-Lite, and Eagle2.5-8B VLM.
OpenAI announced that it will deprecate GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini in ChatGPT on February 13, 2026. This action is synchronized with the previously announced deprecation of the GPT-5 series (Instant, Thinking, Pro); there are currently no changes to the API interface.
Hugging Face has launched the open-source Python library Daggr, used to build AI workflows by connecting Gradio Spaces, Hugging Face Inference Services, and custom functions.
The Tongyi Qwen team has open-sourced the Qwen3-ASR-1.7B and Qwen3-ASR-0.6B speech recognition models, along with the Qwen3-ForcedAligner-0.6B forced alignment model, supporting recognition of 52 languages and dialects.
Revolut has deployed ElevenLabs Agents as the first-tier voice customer service in the UK and Europe, covering over 4 million customers and supporting more than 30 languages. In the initial phase, issue resolution time was reduced by more than 8 times, with over 99.7% of calls successfully connected. This solution retains Revolut's control over business logic and meets PCI compliance and zero data retention requirements.
Google Research published the paper "Towards a science of scaling agent systems," featuring controlled evaluations of 180 agent configurations.
Hugging Face has released the open-source tool upskill, which supports using large models such as Claude Opus 4.5 to generate agent skills (SKILL.md) for specialized tasks like CUDA kernel development. It automatically builds test cases and quantitatively evaluates the impact of these skills on accuracy and token consumption when run locally or with open-source models (such as unsloth/GLM-4.7-Flash-GGUF).
In a paper accepted at ICLR 2026, Google Research proposed ATLAS (Adaptive Transfer Scaling Laws), based on 774 training experiments.
Mistral AI has released Mistral Vibe 2.0, a terminal-native coding agent based on the Devstral 2 model family. New features include custom sub-agents, multi-option clarification, slash command skills, unified agent mode, and automatic updates. It is currently available for Le Chat Pro and Team plan users, supporting pay-as-you-go or bring-your-own API key options. The Devstral 2 models have transitioned to paid API access, while experimental plans continue to offer free quotas.
Hugging Face released the second blog post in its series, analyzing the architectural evolution of China's open-source AI ecosystem since the 'DeepSeek Moment' in January 2025: MoE has become the mainstream choice (e.g., Kimi K2).
Hugging Face, in collaboration with the TII UAE team, released the Alyah benchmark, focusing on the Emirati dialect. It includes 1173 multiple-choice samples manually collected by native speakers, covering 7 categories of phenomena: greetings, religious sensitivity, metaphorical expressions, etiquette values, poetry creation, historical heritage, and dialectal linguistics. The evaluation of 54 Arabic and multilingual large language models found that instruction fine-tuning significantly improves cultural-pragmatic understanding, but dialectal linguistics and everyday expressions remain common challenges.
Teams from LinkedIn and Hugging Face jointly reviewed the technical challenges and engineering fixes in conducting Agentic Reinforcement Learning (Agentic RL) training on the GPT-OSS model: they identified and fixed the issue of PPO's on-policy failure under the MoE architecture.
Figure has released Helix 02, the first end-to-end autonomous system achieving output from pixel input to full-body joint actuation. It supports a 4-minute continuous dishwasher loading/unloading task without human intervention. The system consists of System 0 (a 10M-parameter whole-body motion prior model running at 1kHz), System 1 (a visuomotor policy with full sensor access), and System 2 (a semantic reasoning layer). Leveraging new palm cameras and fingertip tactile sensors on Figure 03 hardware, it completes high-precision operations such as pill extraction and precise syringe injection.
Moonshot AI released Kimi K2.5, claiming it to be the strongest open-source model to date, based on continuous pre-training with approximately 15T mixed image-text tokens, natively supporting multimodality; supports self-directed agent swarms.
MiniMax released MiniMax-M2-her, a model designed specifically for deep role-playing, proposing a modeling framework based on the coordinates of {World} × {Stories} × {User Preferences}.
Moonshot AI released Kimi K2.5, claiming it is the strongest open-source model to date. It is based on continuous pretraining with approximately 15T mixed image-text tokens, natively supports multimodality, and features a self-directed agent swarm supporting up to 100 sub-agents.
The Tongyi Qianwen team officially launched its flagship reasoning model, Qwen3-Max-Thinking, with significantly increased parameter scale and reinforcement learning compute power. Its performance rivals GPT-5.2-Thinking across 19 authoritative benchmarks.
SGLang v0.5.8 has been officially released. Key updates include: up to 1.5x faster inference speed for all mainstream diffusion models; support for chunked pipeline parallelism for ultra-long million-token contexts; and optimized production deployment for GLM4-MoE.
On January 22, Baidu released and launched the official version of its native full-modal large model, ERNIE 5.0. With 2.4 trillion parameters, it employs a unified autoregressive architecture to achieve joint training and native full-modal understanding and generation across text, images, audio, and video. In comprehensive evaluations across more than 40 authoritative benchmarks, its language and multimodal understanding capabilities surpassed models such as Gemini-2.5-Pro and GPT-5-High. It is currently open for enterprise and developer access via the Baidu Qianfan platform.
Google Research published the paper "Small Models, Big Results: Achieving Superior Intent Extraction Through Decomposition".
Moonshot AI open-sourced the Kimi Vendor Verifier (KVV) project alongside the release of the Kimi K2.6 model. It provides six benchmark tests—Pre-validation, OCRBench, MMMU Pro, AIME2025, K2VV ToolCall, and SWE-Bench—to verify the accuracy of open-source model inference implementations. The tool supports execution on NVIDIA H20 servers and includes mechanisms for streaming inference, automatic retries, and checkpoint resumption.