Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
Hugging Face and SkyPilot jointly launched native support for Hugging Face Storage, allowing users to mount Hub model repositories using hf:// URLs and existing HF_TOKENs.
Meta Superintelligence Labs has released Muse Image—the first media generation model with agentic capabilities—supporting precise editing and multi-reference image synthesis.
LeRobot v0.6.0 released, focusing on closing the robot learning loop, introducing three types of world model policies (VLA-JEPA, LingBot-VA, FastWAM), and six new VLAs (GR00T N1.7.
Doubao (ByteDance) released EdgeBench, a benchmark for measuring real-world environment learning. It includes 134 tasks spanning six major domains, with each task supporting at least 12 hours of continuous interaction. Based on approximately 38000 hours of interaction data, it was found that agent environment learning performance highly conforms to a log-sigmoid curve (average R²=0.998), and the environment learning speed of frontier models doubles approximately every three months. 51 tasks, the complete evaluation framework, paper, code, and dataset have been open-sourced.
Doubao (ByteDance Seed) has released EdgeBench, an ultra-long-horizon benchmark designed to measure real-world environment learning. It comprises 134 real tasks covering six major capability domains, with each task supporting at least 12 hours of continuous execution. Based on approximately 38,000 hours of agent-environment interaction data, the study found that overall performance follows a log-sigmoid curve with an average R² of 0.998, and that frontier models' environment learning speed doubles roughly every three months. Currently, 51 tasks along with the complete evaluation framework, paper, code, and dataset have been open-sourced.
The Hugging Face official blog announced a major update to the 🤗 Kernels project, including the addition of a 'kernel' repository type on the Hub to improve discoverability and hardware compatibility querying capabilities.
Immich officially released version v3.0.0, its first major update, which includes several important features: mobile non-destructive image editing (consistent with the web version), workflow automation (drag-and-drop trigger-filter-action orchestration supporting both JSON and visual modes), HLS real-time video transcoding (preview), a new web video player, direct album uploads on Android, OCR text recognition, integrity checks (detecting missing/untracked/checksum-mismatched files), etc. Additionally, support for pgvecto.rs has been removed, and numerous API compatibility changes have been introduced; please refer to the migration guide.
Mistral released Leanstral 1.5, an Apache-2.0 open-source model with 6B active parameters, which reached saturation on miniF2F and solved 587/672 problems on PutnamBench.
Hugging Face and Cerebras have collaborated to launch an open-source, modular real-time voice AI pipeline. It supports the workflow: voice input → Parakeet speech recognition → Gemma 4 31B multimodal inference → Qwen3TTS text-to-speech → voice output. This architecture has been deployed in over 9,000 Reachy Mini robots, with a focus on optimizing P95 latency and multi-turn interaction stability.
IBM Research and Hugging Face released ScarfBench, an open-source AI agent benchmark designed to evaluate framework migration capabilities across three major Java ecosystems: Spring, Jakarta EE, and Quarkus. The benchmark requires that migrated applications must successfully build, deploy correctly, and pass behavioral verification. It comprises 34 applications, 102 framework implementations, and 204 migration tasks. Currently, state-of-the-art agents generally achieve behavioral success rates below 10%.
Google Research published the paper "Estimating high-resolution albedo for urban applications" in Nature Communications.
Google DeepMind has released Nano Banana 2 Lite (gemini-3.1-flash-lite-image), which is the fastest in the Nano Banana series.
ElevenLabs has launched the Procedures feature in ElevenAgents (Alpha stage), supporting both structured and free-form process definitions to guide AI agents in handling tasks such as refunds, billing, and product troubleshooting; it also supports importing from existing SOP documents (PDF/txt, etc.) to automatically generate an initial Procedure.
Google Research has released TabFM, a zero-shot foundation model designed for tabular data classification and regression. By modeling tasks as in-context learning (ICL) problems, it eliminates the need for manual feature engineering, hyperparameter tuning, and model retraining.
Figure officially announced that the Figure 03 humanoid robot has arrived at Plant 52 of the BMW Group's Spartanburg facility to take over from Figure 02 and perform new logistics "sequencing" tasks of increased complexity. These tasks require combining visual perception, dual-arm robotic manipulation, whole-body coordination, and autonomous mobility. Driven by Figure's proprietary VLA model Helix 02, the system achieves real-time motion correction for changes in the positions of unstructured parts.
Hugging Face has launched a deep integration between Every Eval Ever (EEE) and Community Evals: a new YAML converter is added, supporting the automatic mapping of EEE JSON-format evaluation results to the .eval_results/*.yaml files required for Hugging Face model pages.
vLLM v0.24.0 has been officially released, adding support for MiniMax-M3 and DiffusionGemma models, and integrating the Streaming Parser Engine to unify tool calling and reasoning parsing.
NVIDIA researchers proposed the ENPIRE framework, an agentic closed-loop system for self-improving robot policies in the real world. It comprises four core modules: Environment Reset (EN), Policy Improvement (PI), Policy Rollout (R), and Evolution (E). The system achieved success rates up to 99% on tasks including PushT, organizing a pin box, cutting zip ties, and inserting GPUs into motherboards. Each workstation is equipped with two I2RT YAM robotic arms and an NVIDIA RTX 5090.
Researchers accidentally discovered during laboratory tests in 2024 that ordinary 180-nanometer CMOS transistors with unconnected bulk terminals can exhibit biological neuron-like spiking and hysteresis characteristics; hole accumulation in the body region triggers the conduction of a hidden bipolar junction transistor. Furthermore, it was found that by controlling the bulk terminal voltage, the same transistor can controllably trap charges in the gate dielectric, achieving stable and tunable synaptic conductance. This phenomenon is 100% reproducible on chips manufactured by different wafer fabs, with device lifetimes exceeding 10 million cycles.
Meta released Brain2Qwerty v2, an end-to-end non-invasive brain signal decoding model that can directly convert MEG-recorded brain activity into sentences in real-time scenarios, achieving a word accuracy of 61%, with the best participant reaching 78%; simultaneously, all training code for v1/v2 and the v1 dataset were open-sourced by BCBL.
SGLang v0.5.14 has been officially released, adding support for GLM-5.2, LiquidAI LFM2.5, Kimi-K2.7-Code, and Poolside Laguna-M.1.
Google Research proposes a frozen Multi-Token Prediction (MTP) architecture that attaches lightweight MTP heads to the deployed Gemini Nano v3 model without modifying weights, achieving on-device inference acceleration. The approach uses a zero-copy design, reuses the main model's KV cache, saves 130MB of memory, and achieves up to a 55% improvement in token acceptance rate and over 50% task-level speedup on Pixel 9 devices.
The official Hugging Face blog announced that private, OpenAI-compatible vLLM services can be launched via a single command on HF Jobs. It supports models such as Qwen/Qwen3-4B, automatically exposes ports, and provides HTTPS access URLs; it also supports hardware specifications like a10g-large or h200x2, SSH debugging, Gradio UI integration, and backend integration for the Pi coding agent.
SungMo Park, a researcher at a16z crypto, analyzes how South Korea is shifting from a retail crypto market to institutional-grade blockchain infrastructure development, with a core focus on KRW stablecoins and real-world asset (RWA) tokenization.