↗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

AI products · Hugging Face Blog

llama.cpp Releases Router Mode for Dynamic Multi-Model Loading and Switching ↗

The llama.cpp server has added a router mode that supports dynamically loading, unloading, and switching multiple GGUF models without restarting the service. It ensures stability through multi-process isolation and offers capabilities such as auto-discovery, LRU unloading, on-demand loading, and model routing, with support for CLI parameters, preset configurations, and Web UI switching.

Practical tips · Hugging Face Blog

Hugging Face Releases HF-skills Training Plugin, Supporting End-to-End Model Fine-Tuning with Codex ↗

Hugging Face has released the open-source HF-skills plugin, enabling OpenAI Codex to perform end-to-end machine learning experiments: validating dataset formats, automatically selecting hardware (e.g., t4-small), generating and submitting training scripts, monitoring progress via Trackio, evaluating on openai_humaneval, generating training reports, exporting GGUF quantized models, and publishing them to the Hub. It supports training methods such as SFT, DPO, and RLHF, and is applicable to models with 0.5B–7B parameters.

AI models · Runway Research

Runway Releases General World Model GWM-1 ↗

Runway has released GWM-1, its first General World Model family, built on Gen-4.5 and supporting action conditioning (e.g., camera poses, robot commands).

AI models · 智谱官方研究

Zhipu Releases Industrial-Grade Speech Synthesis Model GLM-TTS ↗

Zhipu officially released the speech synthesis model GLM-TTS, with open-source weights available on Hugging Face and ModelScope; the model is based on a two-stage generation paradigm (Text-to-token Auto-regression + Token-to-wav Diffusion).

AI products · SGLang Releases

SGLang Model Gateway v0.2.4 Released ↗

SGLang officially released Model Gateway v0.2.4, focusing on enhanced performance, security, and production-grade observability: optimized radix tree implementation for cache-aware load balancing, updated Tokenizer and core modules (HTTP/gRPC routers), introduced the industry's first WASM middleware support, and completed full-chain OpenTelemetry integration (including HTTP/gRPC distributed tracing).

AI models · 智谱官方研究

Zhipu Releases Open-Source Speech Recognition Model GLM-ASR-Nano-2512 ↗

Zhipu has released the open-source speech recognition model GLM-ASR-Nano-2512 with 1.5 billion parameters, achieving an average error rate of 4.10 among open-source models, outperforming Whisper V3 (6.93). It supports Cantonese, mixed Chinese-English, low volume, noisy environments, and hotword recognition, and has been integrated into Zhipu AI's input method.

Research · Google DeepMind

Google DeepMind Releases FACTS Benchmark Suite to Systematically Evaluate LLM Factuality ↗

Google DeepMind, in collaboration with Kaggle, released the FACTS benchmark suite, which includes four sub-benchmarks: Parametric, Search, Multimodal, and the updated Grounding v2, totaling 3,513 public examples. The FACTS Score is the average accuracy across these four benchmarks, with Gemini 3 Pro leading at 68.8%. All datasets include public/private splits, and Kaggle hosts the leaderboard.

Research · Google DeepMind

Scientists Use AlphaFold to Analyze the Structure of Plant Photosynthetic Enzyme GLYK and Design Heat-Resistant Variants ↗

A team led by Berkley Walker at Michigan State University used AlphaFold to predict the 3D structure of GLYK, a key photosynthetic enzyme in plants and thermophilic algae. They found that plant GLYK becomes unstable at high temperatures due to three flexible loops; after validation through molecular simulations, they replaced these loops with rigid ones from the algal version, creating a hybrid enzyme that remains stable at 65°C.

AI models · 通义／Qwen 官方研究

Tongyi Qwen3-TTS Released: A Speech Synthesis Model Supporting 49 Timbres, 10 Languages, and 9 Dialects ↗

Tongyi Qwen3-TTS is a newly released flagship speech synthesis model that supports 49 high-quality timbres, 10 major languages (including Chinese, English, Japanese, etc.), and 9 Chinese dialects (such as Cantonese, Sichuanese, Beijing dialect, etc.). It achieves an average word error rate superior to MiniMax on the MiniMax TTS multilingual test set.

Practical tips · Hugging Face Blog

Hugging Face Releases hf-llm-trainer Skill Package, Enabling Full Automatic Fine-Tuning of Open Source Large Language Models by Coding Agents like Claude ↗

Hugging Face has launched the hf-llm-trainer skill package, enabling coding agents such as Claude Code, Codex, and Gemini CLI to fully automatically complete model fine-tuning: validating dataset formats, selecting GPUs (e.g., t4-small), generating and submitting training scripts, monitoring progress, and pushing models to Hugging Face Hub.

AI products · Hugging Face Blog

Hugging Face Transformers v5 Released ↗

Hugging Face released Transformers v5.0.0rc-0, with the core goal of improving interoperability, focusing on refactoring the model definition approach (introducing AttentionInterface.