AI Daily — 2026-10-14

English 中文

GreenWAM introduced a robotics world model, Tev scaled long-horizon agentic environments with dense rewards, and NVIDIA will release Nemotron Bolt 3.5

Covering 21 AI news items

🔥 Top Stories

1. NVIDIA Nemotron Bolt 3.5 74B A9B Releases October 15

NVIDIA’s upcoming 74B-parameter MoE model uses only 9B active parameters and 2-bit quantization to fit within 24GB, with quantization-aware training to preserve quality. The release pushes high-capability local inference closer to consumer hardware, while the smaller E4B 4GB variant on October 23 signals a broader push into edge deployment. Source-reddit

2. China’s Open-Weight AI Models Trail US Closed Models by 7 Months

An open leaderboard aggregating 22,113 benchmark results finds Chinese open-weight models roughly seven months behind US closed frontier systems. The split between independent and lab-reported scores highlights both progress and measurement risk, giving enterprises a clearer—if still contested—map of the open-weight frontier. Source-reddit

3. inclusionAI Releases Ling-3.1-flash 560B MoE with 1M Context

inclusionAI’s Ling-3.1-flash scales to 560B total parameters with 25B active per token and a 1M-token context window, up sharply from Ling 3.0’s 127B. The release targets long-document, retrieval-heavy, and agentic workloads, though serving such a large MoE will demand substantial memory and routing efficiency. Source-reddit

📰 Featured

Robotics & Embodied AI

GreenWAM: A World Foundation Model for Robotics — GreenWAM proposes a shared anatomical action space and training recipe to serve as both world model and policy across robot bodies, tackling fragmented action formats in robotics. Source-huggingface
Urbex: Agentic Vision-Language Framework for City-Scale 3D Grounding — Urbex reframes city-scale 3D grounding as active spatial search, letting a VLM agent query landmarks, zoom, render oblique views, and commit bounding boxes for more efficient large-scene understanding. Source-huggingface

Agentic AI & Training

Tev Scales Long-Horizon Agentic Environments with Dense Progress Rewards — Tev introduces dense progress rewards to train agents on tasks lasting hours to days, addressing sparse rewards and data bottlenecks that limit open-source agentic models. Source-huggingface
Agent-Reflex: Training Language Agents to Reflect Proactively — Agent-Reflex uses a three-stage training framework to help language agents recover from noisy MCP tool feedback and user interventions through proactive reflection. Source-huggingface

Efficient Open-Weight Models

MiniCPM-V 4.7 Runs 35B Multimodal MoE on Phones — OpenBMB demonstrates a 35B-A3B vision-language MoE running locally on mobile with around 3GB memory, enabled by LES expert scheduling and a smaller 1B variant for on-device multimodal tasks. Source-reddit
Telperion Releases Qwen3.8-27B MLA/GLA Variants with 4x KV Cache Compression — Telperion’s latent-attention variants cut KV cache by 4x, with the MLA model improving SWE-bench Verified to 42.1% while using about 20% fewer reasoning tokens. Source-reddit
Sabi-Noema: Qwen 3.8 27B Fine-Tune Boosts Capability, Efficiency — Sabi-Noema and Sabi-Noema-MLA improve coding, agentic, and tool-calling performance while halving verbosity and reducing hallucinations, reportedly beating GPT 6 Luna-medium on DeepSWE. Source-reddit

⚡ Quick Bites

Open-Source VoiceStudio Offers Local ElevenLabs Alternative for 646 Languages — VoiceStudio provides a local, open-source voice synthesis option spanning 646 languages for privacy-sensitive audio workflows. Source-github
Vercel Labs Launches Open Agent Skills CLI for Coding Agents — Vercel Labs released a CLI for adding reusable agent skills to coding agents. Source-github
Mr.Chatterbox-2.7B-Chat: Victorian Text-Trained Language Model Released — A 2.7B chat model trained from scratch on Victorian-era text has been released for stylistic and historical language experiments. Source-reddit
Zellige AI Releases Tessera-27B-B-CJK Qwen Finetune for Token-Efficient Thinking — Zellige AI’s Qwen-based finetune targets CJK use cases with more token-efficient reasoning. Source-reddit
Anthropic Exec Says Open-Source AI Models Are Higher Risk — An Anthropic executive argued that open-source AI models carry higher risk, reigniting debate over openness and safety. Source-reddit
Zero-Setup Local Photo Editor Runs Qwen-Image 2.1 on 12GB+ GPUs — A zero-setup local photo editor runs Qwen-Image 2.1 on GPUs with 12GB or more VRAM. Source-reddit
ComfyUI v0.40.0 Released — ComfyUI v0.40.0 ships updates for the popular node-based image-generation workflow tool. Source-reddit
Jeva: System One Models from First Principles — Jeva proposes a first-principles approach to “System One” models for fast, intuitive reasoning. Source-huggingface
BreezeBlue-TTS-2 Claimed Best Local TTS Model, Non-Commercial License — BreezeBlue-TTS-2 is being promoted as the best local TTS model, though its non-commercial license limits deployment. Source-reddit
Jailbroken PS5 Runs Qwen 3.8 27B Locally at 20 Tokens/Second — A jailbroken PS5 ran Qwen 3.8 27B locally at about 20 tokens per second, showing unconventional consumer-hardware inference. Source-reddit
User Compares Qwen 3.8 Flash Next vs DeepSeekV4 Flash Speed — A user benchmark compares Qwen 3.8 Flash Next and DeepSeekV4 Flash on speed and latency. Source-reddit

Generated by AI News Agent | 2026-10-14