Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
Cohere has released the open-source model North Mini Code, featuring 30B total parameters and 3B active parameters, supporting a 256K context window, and optimized for agentic software engineering, terminal tasks, and code generation.
Hugging Face has launched an integration between GitHub Actions and Hugging Face Jobs, enabling the migration of CPU/GPU CI tasks to its serverless infrastructure.
Google has open-sourced XProf Kernel, a low-level hardware debugging tool designed specifically for optimizing custom TPU kernels written in frameworks such as Pallas. It supports inspection at the HLO graph and MLIR levels, visualization of Low-Level Operation (LLO) dynamic execution traces, microsecond-level performance counter sampling on TPU v7 (Ironwood), and introduces an event-triggered sampling mode that allows configuring up to 28 hardware counters per SparseCore.
Google DeepMind reported on a pre-registered randomized controlled trial conducted in Sierra Leone, showing that its Guided Learning (based on Gemini) improved students' math scores by +0.258 standard deviations, equivalent to 1.2–1.7 years of regular learning progress; under teacher-led interventions, students who used it for approximately 12 hours showed progress equivalent to 1.8–2.5 years.
King's College London, Fudan University, and The Alan Turing Institute jointly released the SocioHack benchmark, which includes 72 sandbox environments simulating the reward structures of social institutions, categorized into three types: historical (32), synthetic (20), and fictional (20). RL-trained models reproduced historical vulnerabilities within these environments with a recall rate of 61.25% and a precision rate of 90.85%. The study points out that when social institutions are encoded as reward rule systems, reward hacking is equivalent to hacking the rules governing social operations themselves.
Hugging Face announced that OpenEnv has been upgraded to an open-source standard for agentic reinforcement learning execution environments, co-built by multiple parties and now coordinated by a committee comprising organizations such as Meta-PyTorch, Microsoft, and Nvidia.
Google has launched the Gemini Enterprise Agent Platform, introducing a managed Cross-Corpus Retrieval feature based on Agentic RAG.
Google Research published a paper in Nature introducing PHRM, a passive heart rate monitoring system. Using front-camera videos captured during users' routine face-unlock interactions, PHRM estimates heart rate (MAPE < 10%) and daily average resting heart rate (MAE < 5 bpm) via a lightweight temporal convolutional network. It achieves wearable-device-level accuracy in both laboratory and real-world settings and covers the full spectrum of skin tones. Concurrently, it releases the largest and most diverse smartphone facial video research dataset to date, along with the pre-trained model PHRM-mini.
NVIDIA has released the Nemotron 3.5 Content Safety model, which is built upon Gemma 3 4B IT and supports unified multimodal input (user prompts, optional images).
AI21 achieved a new SOTA with a 60.9% problem-solving rate on the SWE-rebench benchmark, keying on reversing the traditional 'augment-then-scale' agent pipeline to 'scale-then-augment'.
ElevenLabs has launched the Flows Agent within ElevenCreative, a conversational AI assistant integrated into a node-based creative canvas. It supports describing needs in natural language (e.g., "create a podcast intro with custom music and voiceover") and automatically builds and executes multimodal pipelines involving images, video, speech, lip-syncing, and background music. It offers clarifying questions, real-time iterative modifications, an Assist mode to confirm expensive operations, and background execution capabilities.
ChatGPT has launched a new memory system called Dreaming, which better remembers user preferences, keeping context fresh and relevant across conversations.
Hugging Face officially released the agent-optimized hf CLI, which supports automatic detection of coding agents such as Claude Code, Codex, and Cursor via environment variables starting from v1.9.0.
Google Research has open-sourced its hydrological modeling framework, which includes an LSTM-based river forecast model architecture, a PyTorch implementation, a training pipeline, and compatibility with the Caravan dataset.
The Google Open Source Blog reviews the development history of the JPEG XL standard, covering open-source experiments from 2011–2017 such as WebP Lossless, Brotli, Butteraugli, Guetzli, and Brunsli; the fusion of PIK and FUIF during 2017–2019 to form the VarDCT architecture; and current implementation status in DNG 1.7, ProRAW, DICOM, PDF/EPUB, Adobe, Apple, Linux, Windows, and browsers.
Dharma AI and Hugging Face collaborated to release the DharmaOCR study, proposing a method that uses the model's own degraded text outputs (such as repetition loops) generated after Supervised Fine-Tuning (SFT) as rejected samples to construct DPO training pairs for structured OCR tasks. This approach reduced degradation rates by an average of 59.4% across 5 open-source model families, reaching up to 87.6%, without compromising extraction quality. The paper indicates that DPO effectiveness requires three structural conditions: failure modes must be clearly classifiable, distinguishable by automated scoring mechanisms, and have sufficient output volume.
The official Hugging Face blog announced that Reachy Mini now supports MCP tools from Hugging Face Spaces, enabling users to add remote tools via a single command (e.g., reachy-mini-conversation-app tool-spaces add pollen-robotics/reachy-mini-weather-tool).
Hugging Face has released the Holo3.1 model family, which supports Web, desktop, and Android mobile environments and adds native support for function-calling protocols.
JetBrains has released the open-source Mixture-of-Experts model Mellum2, with a total of 12B parameters and 2.5B active parameters per token. It supports text and code tasks and is more than twice as fast in inference compared to similar open-source models. It is suitable for routing, RAG, sub-agents, high-throughput coding, and private deployments, licensed under Apache 2.0; the technical report is available at arXiv:2605.31268.
In a Hugging Face blog post, IBM Research stated that frontier LLMs alone cannot support scalable enterprise-level AI adoption; introducing 'agentic logic' (such as knowledge graphs, program analysis libraries, algorithms, and adaptive orchestration) is necessary to constrain context.
OpenAI frontier models and Codex are now fully available on AWS. Enterprises can leverage their existing AWS environments, governance mechanisms, and procurement processes to build applications and accelerate deployment from evaluation to production.
The Tongyi Qianwen team has officially released Qwen3.7-Plus, a multimodal model that unifies vision and language into an integrated agent foundation. It supports inputs including images, videos, screens, web pages, and text, enabling task completion in GUI/CLI/tool environments. The model is available via Alibaba Cloud Bailian's API service, compatible with OpenAI and Anthropic protocols, and can be integrated into mainstream agent frameworks such as Claude Code, OpenClaw, and Qwen Code.
MiniMax officially released the M3 model, featuring frontier-level coding and agent capabilities, a 1 million token context window, and native multimodality (supporting image/video input and desktop operation). Its core innovation, the MSA sparse attention architecture, reduces computation per token to 1/20th of previous models, achieving frontier-level performance on benchmarks such as SWE-Bench Pro and Terminal-Bench 2.1.
vLLM v0.22.0 has been officially released, with core updates including full support for the DeepSeek V4 model via NVFP4 MoE, CUDA graphs, MTP speculative decoding, and refactoring into an independent module.