Several chip manufacturers are advancing Complementary Field-Effect Transistor (CFET) technology, which vertically stacks p-type and n-type nanosheet transistors to enhance integration density. Intel, Samsung, and TSMC adopt monolithic processes, while IBM primarily promotes its sequential process, Nanostack, claiming it can achieve a 50% performance boost, a 70% improvement in energy efficiency, and a 40% reduction in memory area compared to 2nm node chips.
Google Research will publish the paper "Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs" at COLM 2026, revealing that Chain-of-Thought (CoT) reasoning significantly improves answer recall rates even in single-hop factual question answering.
Google DeepMind announced that computer use capabilities have been integrated as built-in tools into the Gemini 3.5 Flash model, no longer limited to the previously standalone Gemini 2.5 computer use model.
A research team at Princeton University proposed an end-to-end AI-driven radio frequency integrated circuit (RFIC) design method. By combining a reinforcement learning framework with electromagnetic structure inverse design based on diffusion models, it generates chip layouts for broadband millimeter-wave power amplifiers without human templates, achieving state-of-the-art levels in bandwidth, output power, and efficiency among silicon-based chips at the time. This method reduces the design cycle from months to minutes and supports multi-port structures and interpretable control.
Mistral AI has introduced new Connectors features, including fine-grained permission management for workspaces and organizations (GA), scoped API keys (GA), multi-account authentication connectors (GA), Connectors Debugger (public preview), connector integration in Workflows (public preview), and connector reuse in Vibe Code (GA).
On the DeepResearch Bench II (DRB II) benchmark, AI21 achieved a total score of 64.38 by aggregating outputs from weak agents ranked 7–13 (each reporting total scores of 38–45) using pairwise aggregation with Gemini 3.1 Pro, surpassing the previous SOTA by 3.2 points; this method does not rely on new models or training, with its core being the enhancement of Information Recall, and it validates an adjustable trade-off between the quantity and quality of aggregation.
OpenAI and Broadcom have launched a custom AI chip named Jalapeño, designed specifically for large language model inference, aiming to improve performance, energy efficiency, and system scalability.
Hugging Face and Treble Technologies jointly released the FFASR Leaderboard, the first open-source, community-driven benchmark for far-field automatic speech recognition (ASR).
Mistral AI released Mistral OCR 4, a small model focused on document understanding that supports 170 languages and returns structured output with bounding boxes, block type classification (e.g., headings/tables/formulas), and word/page-level confidence scores.
Tongyi Lab officially released Qwen-AgentWorld, the first native language world model (LWM), covering MCP, Search, Terminal, SWE, Web, and OS in a single model.
A guest post written by engineers from the Google Chrome team was published on the Hugging Face blog, introducing how to experimentally use the Cross-Origin Storage (COS) API, which is still in the proposal stage, within Transformers.js.
Hugging Face has open-sourced the automated weekly release pipeline for huggingface_hub. Orchestrated by GitHub Actions, it uses an OpenCode Agent running the GLM-5.2 model to draft release notes and Slack notifications. Accuracy is ensured through deterministic PR checklist validation and human review. The entire process relies on open-source tools, open-weight models, and OIDC trusted publishing, with all workflow files made public.
PaddleOCR has released PP-OCRv6, which includes three model tiers: tiny (1.5M parameters), small (7.7M parameters), and medium (34.5M parameters). The medium version achieves 86.2% detection Hmean and 83.2% recognition accuracy on a self-built multi-scenario benchmark.
Researchers from the University of Oxford, the UK's AI Safety Institute (AISI), Stanford University, and the London School of Economics found through four experiments (totaling 18,978 conversations and 6,923 participants) that current AI systems [surpass human experts] in real-world scenarios such as policy issue persuasion and charitable fundraising.
ElevenLabs officially announced the launch of Ads Engine within ElevenCreative. It supports connections to Google Ads, Meta Ads, and LinkedIn ad platforms, offering advertising localization capabilities including text translation, image adaptation, and video dubbing (Dubbing V2). The feature also supports template reuse and ad fatigue detection. This functionality is now live, accessible at elevenlabs.io/ads-engine.
Kubernetes v1.35 introduces the RestartAllContainers action, supporting in-place restart of all containers without destroying the Pod object; this feature is enabled via the RestartAllContainersOnContainerExits feature gate, upgraded to Beta and enabled by default in v1.36.
Hugging Face, in collaboration with researchers, released MosaicLeaks, a new benchmark targeting privacy leakage in deep research agents, focusing on mosaic-style leaks caused by the interweaving of public and private information in multi-hop questions; it proposes the Privacy-Aware Deep Research (PA-DR) reinforcement learning method, which reduces the full-information leakage rate from 34.0% to 9.9% while maintaining a strict chain success rate of 58.7%; the paper has been submitted to arXiv (2605.30727).
Hugging Face has released an open-source evaluation framework called agent-eval, designed to measure the actual performance of agents driven by open-source models when using specific tools (such as transformers).
In its official blog, Hugging Face released a systematic empirical benchmark of LoRA and over 40 other Parameter-Efficient Fine-Tuning (PEFT) techniques, covering two major tasks: LLM mathematical reasoning (MetaMathQA/GSM8K) and image generation (learning the plush cat concept using FLUX.2-klein-base-4B).
AWS's open-source Strands Robots SDK (Apache 2.0) has released LeRobot integration, encapsulating LeRobot datasets, policies, and hardware abstractions as AgentTools.
Zhipu released the GLM-5.2 large language model, achieving stable and usable 1M token context in an open-source model for the first time, specifically designed for long-horizon tasks; the new model adopts the IndexShare architecture to reduce FLOPs and improves the MTP layer to increase speculative decoding acceptance length by 20%.