↗TiaoTechIdeas in motion
中文
☰
Home/Stay curious. Stay in the loop.
TIAOTECH / NEWS

Stay curious. Stay in the loop.

Selected industry news from AIHOT. Follow advances in AI, products and technology.

AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM

Practical tips · Hugging Face Blog

Comprehensive Support for Mixture-of-Experts (MoE) Models in Hugging Face Transformers ↗

A Hugging Face blog post details the engineering implementation of Mixture-of-Experts (MoE) models within the Transformer architecture, covering weight loading reconstruction (supporting expert tensor packing and dynamic conversion) and expert backends (three execution strategies: eager, batched_mm, and grouped_mm).

AI products · vLLM Releases

vLLM v0.16.0 Released ↗

vLLM has released version v0.16.0, featuring core updates including asynchronous scheduling and pipeline parallelism support (30.8% end-to-end throughput improvement), a WebSocket real-time API, improved RLHF workflows, unified parallel speculative decoding, and XPU platform kernel refactoring. It also adds support for model architectures such as GLM-OCR, Qwen3-ASR, DeepSeek-OCR-2, Intern-S1-Pro, MiniCPM-o 4.5, and GLM-5.

Industry · Ethereum Foundation Blog

Ethereum Foundation Launches Treasury Staking Program ↗

The Ethereum Foundation has launched a treasury staking program, solo staking approximately 70,000 ETH with rewards flowing back to the foundation's treasury. It utilizes open-source software Dirk (distributed signing) and Vouch (multi-client support), employs Type 2 (0x02) withdrawal credentials, and is deployed on a hybrid infrastructure combining multi-location hosting and self-managed hardware. Deposits for the initial batch of validators have been made public, with the remainder expected to be completed within weeks.

AI models · 智谱官方研究

Zhipu Releases GLM-5 Large Language Model ↗

Zhipu has launched its next-generation foundation model, GLM-5, focusing on Agent, Reasoning, and Coding (ARC) capabilities. It adopts DeepSeek Sparse Attention (DSA) to reduce inference costs, builds an asynchronous reinforcement learning infrastructure, and natively supports seven major domestic chip platforms including Huawei Ascend.

Industry · Hugging Face Blog

GGML and llama.cpp Team Join Hugging Face to Advance Local AI Development ↗

Hugging Face announced that the GGML team (creators of llama.cpp) has officially joined, with the goal of scaling support for the local AI community; the llama.cpp project remains 100% open source, technically autonomous, and community-driven, with Hugging Face providing long-term sustainable resources; both parties will promote seamless integration of model definitions between llama.cpp and the transformers library, and optimize the packaging and user experience of ggml software to make local inference easier to deploy.

AI models · Google DeepMind

Google DeepMind Releases Gemini 3.1 Pro Model ↗

Google DeepMind has released the Gemini 3.1 Pro model, describing it as a more intelligent baseline model for the most complex tasks, achieving a 77.1% validation score on the ARC-AGI-2 benchmark.

AI products · Google DeepMind

Gemini App Launches Lyria 3 Music Generation Model (Beta) ↗

On February 18, 2026, the Gemini app launched Lyria 3 in beta, enabling the generation of 30-second music from text or images. Features include automatic lyric composition, control over style/vocals/rhythm, and improved audio quality and complexity. All generated audio embeds SynthID watermarks, and a new AI verification feature for audio content has been added. Currently supports 8 languages; desktop rollout is complete, with mobile rollout expected within days.

Industry · Ethereum Foundation Blog

Ethereum Foundation Releases 2026 Protocol Priorities: Establishing Three New Tracks - Scale, Improve UX, and Harden the L1 ↗

The Ethereum Foundation released an update on its 2026 protocol priorities, announcing that the previous three-track structure (Scale L1/Scale Blobs/Improve UX) will be upgraded to three new tracks: Scale (focusing on consensus, execution, and Blob scaling), Improve UX (focusing on native account abstraction and interoperability), and Harden the L1 (focusing on security, censorship resistance, and network resilience). It reviewed the results of the two mainnet upgrades in 2025, Pectra and Fusaka, including EIP-7702, PeerDAS launch, gas limit increase to 60M, and historical data expiry. It clarified that the Glamsterdam upgrade will land in the first half of 2026.

Research · MiniMax 官方新闻

MiniMax Releases Forge, a Scalable Agentic Reinforcement Learning Framework ↗

MiniMax has released the Forge agentic reinforcement learning framework. Through a flexible system architecture, the CISPO algorithm adapted for long-horizon agents, asynchronous scheduling, and training/inference optimizations, it breaks through the trilemma conflict between system throughput, training stability, and agent flexibility. This framework supports MiniMax M2.5 training, handling over 100,000 types of real-world agent scaffolds, with context lengths up to 200k and daily sample throughput reaching the million level.

AI products · OpenAI News

OpenAI Releases Open-Source Tool GABRIEL ↗

OpenAI has released the open-source tool GABRIEL, which uses GPT to convert qualitative text and images into quantitative data, helping social scientists scale their research analysis.

AI models · MiniMax 官方新闻

MiniMax Releases M2.5 Large Language Model ↗

MiniMax released the new model MiniMax-M2.5, stating that it reaches frontier-level performance in coding (80.2% on SWE-Bench Verified), agent tool calling, search (76.3% on BrowseComp), and office tasks.

Practical tips · Hugging Face Blog

OpenEnv in Practice: Evaluating Tool-Calling Agents in Real Environments ↗

Meta and Hugging Face jointly launched the open-source framework OpenEnv to evaluate AI agents in real systems (rather than simulated environments); Turing contributed Calendar Gym, a production-grade calendar management environment that supports realistic constraints such as access control, temporal reasoning, and multi-agent coordination.