Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
SGLang v0.5.15 has been officially released, with a key focus on optimizing the production inference performance of GLM-5.2 NVFP4 on the Blackwell architecture (achieving over 500 tokens per second per user on an 8x B300 setup).
The GitHub Engineering team replaced Copilot's originally self-developed code exploration tools with the grep, glob, and view tools shared by Copilot CLI.
The official Hugging Face blog published the "Profiling in PyTorch (Part 3)" tutorial, systematically comparing naive attention with its in-place variants.
Sakana AI, in collaboration with MIT and NYU, published the paper "In Search of the Ingredients of Open-Endedness: Replicating Picbreeder with Large Vision-Language Models" at GECCO 2026, using VLM agents to reproduce a goal-free image co-evolution experiment.
Analysis by Marketplace Pulse shows that as of July 2026, Chinese sellers accounted for 55.9% of the top 10,000 sellers on Amazon US, an increase of 1,342 seats compared to July 2020. The share of US sellers dropped to 40.5%, a decrease of 1,320 seats.
Mistral Studio has launched systematic governance features for Prompts and Skills, providing version control, clear ownership, audit logs, classification tags, and quick rollback capabilities. It supports direct editing and testing by non-developers and deploys to production environments via existing CI/CD pipelines triggered by tags. All assets are managed centrally within Studio and integrated with runtime observability, ensuring that only controlled versions are executed.
ChatGPT Work is an agent that can perform actions within users' applications and files, supports long-term project follow-up, and converts goals into completed work.
Google Research released SensorFM, a large sensor foundation model pre-trained on trillions of minutes of unlabeled wearable sensor data from 5 million participants; it employs an Adaptive and Inherited Masking (AIM) self-supervised framework to handle fragmented data; it outperforms feature engineering baselines on 35 health tasks, and its predictions can substitute for real measurements to support personal health agents in generating clinical-grade summaries.
The Ethereum Foundation's Protocol Security team shared practices for using coordinated AI agents to conduct security audits on Ethereum protocol code (such as libp2p gossipsub), discovering a real vulnerability CVE-2026-34219. They emphasized that agents are essentially search tools rather than oracles, with the key challenge lying in classification and filtering (triage)—distinguishing true findings from high-confidence noise through reproducible PoCs, clear invariants, independent verification, and deduplication mechanisms. The team adopted a four-role collaborative pipeline consisting of recon, hunting, gap-filling, and validation. All candidates must satisfy six elements: objective, invariant, mechanism, observable proof, runnable reproducer, and deduplication key, with final disclosure decisions made by humans.
Meta has released Muse Spark 1.1, a multimodal reasoning model designed for agentic tasks. It supports a million-token context window and zero-shot generalization for invoking native tools, MCP servers, and custom skills. The model significantly outperforms its predecessor in computer use, coding, and multimodal understanding. Concurrently, Meta opened the public preview of the Meta Model API; the model is now available in the Meta AI app and via the 'Thinking' mode on the meta.ai website.
PyTorch 2.13.0 has been officially released. Major updates include: FlexAttention now supports Apple Silicon (MPS) and adds a deterministic backward path for CUDA; CuTeDSL serves as a new high-performance GPU code path for Inductor.
Mistral AI has released Robostral Navigate, an 8-billion-parameter embodied navigation large model that performs instruction navigation tasks in complex environments using only a single RGB camera. It achieved a 76.6% success rate on the R2R-CE unseen set, surpassing the previous best single-camera solution by 9.7 points and the best multi-sensor solution by 4.5 points. The model is fully self-developed, trained purely in simulation, supports wheeled/legged/flying robots, and employs goal-directed navigation and online reinforcement learning (CISPO) to enhance robustness and adaptability.
A team from KAIST in South Korea proposed the V-Die solution, which achieves speeds 82% higher than HBM4, 37% lower read latency, and 540 tokens/s (under GPT-3 load) with 16-layer stacking by integrating microfluidic cooling channels into laterally stacked DRAM, keeping peak temperatures at 45°C; teams from the University of Tokyo and others in Japan proposed the MOSAIC solution, using inductive coupling interconnects to provide 294 GB of memory with 98-layer stacking and a peak temperature of 81.3°C, reaching 294 layers and 882 GB when thinned to 100 µm.
OpenAI releases GPT-Live, a next-generation voice model that now powers ChatGPT Voice.
Hugging Face's official blog announced that the transformers modeling backend for vLLM has achieved inference throughput equal to or faster than the native vLLM implementation. This covers three types of models: Qwen3-4B (single GPU), Qwen3-32B (tensor parallelism), and Qwen3-235B-A22B-FP8 MoE (data + expert parallelism). Users can enable this simply by adding the --model-impl transformers flag, without needing to modify their existing deployment configurations.
Hugging Face and Amazon SageMaker Studio have launched a one-click deep link integration, enabling users to click 'Customize on SageMaker AI' or 'Deploy on SageMaker AI' on Hugging Face model pages to directly access the corresponding workflows in Studio.
Google Research published the paper "Urban congestion relief experiments through routing-app interventions" in *Nature Cities*.
Doubao has officially released the multimodal image creation model Seedream 5.0 Pro. Compared to its predecessor, it features comprehensive improvements in image-text matching, structural rationality, text rendering, and aesthetic appeal. It also introduces four new capabilities: complex information visualization, interactive precision editing, realistic imagery and portrait texture, and native multilingual input and generation. The model is now available on the Volcano Ark Experience Center, Doubao, and Jimeng platforms.
Doubao officially released Seedream 5.0 Pro, a multimodal image generation model that comprehensively improves text-image matching, structural rationality, text rendering, and visual aesthetics. It achieves four major breakthroughs: complex information visualization, interactive precise editing (supporting point selection/circle selection/sketch rendering/layer separation/multi-image fusion), realistic imagery and portrait texture, and native multilingual input and generation. The model is now available on the Volcano Ark Experience Center, Doubao, and Jimeng.
AI21 released a study proposing two budget-aware execution strategies: cascading and parallel. These strategies work by predicting the performance of LLM selectors on candidate pools (Resolve-Now / Resolve-Later) and integrating outputs from test agents.
At Build 2026, Microsoft announced that Hugging Face models are officially integrated into Foundry Managed Compute, offering a curated collection of open-source models updated weekly.
Cohere has released its open-source Arabic speech recognition model, Cohere Transcribe Arabic. Built on its 2B-parameter frontier ASR model, it is designed specifically for commercial and development scenarios involving Arabic, covering dialectal differences, Arabic-English code-switching, and domain-specific terminology.
Berkeley Artificial Intelligence Research (BAIR) published a perspective article titled "Intelligence is Free, Now What? Data Systems for, of, and by Agents".
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.