Trending AI Posts & Hot Research Catchup - Page 153
Alibaba Qwen3.8-2.4T-A95B Explained: Open Weights, 2.4T MoE, and the Real Cost of Self-Hosting
19 sources0Qwen3.8 2.4T A95B is Alibaba’s open weight, text only Qwen Max class model: 2.4 trillion parameters with roughly 95 billion active per token.
China Unicom’s Guangdong Marine Network: How Shore, Sea, Ships and Satellites Work Together
22 sources0China Unicom says its September 8, 2026 Guangdong launch combines nearly 1,500 coastal stations, offshore wind farm sites, mobile shipborne base stations and satellite switching to extend maritime connectivi...
DeepSeek V4.1 Flash: 552B MoE, 1M Context, and What Changed From V4 Pro
22 sources0DeepSeek V4.1 Flash, released September 10, 2026, is a 552B multimodal MoE model that activates about 8B parameters for prompt prefill and 16B for generation.
JD Open-Sources JoyAI-EchoWM, an Interactive AI World Model With Synchronized Sound
17 sources0JD announced open source JoyAI EchoWM and the four step EchoWM Flash at JDD 2026.
HBM Shortages Push Up Chinese AI Chip Prices and Raise 2026 Supply Risks
16 sources0HBM scarcity has made Chinese AI accelerators more expensive and harder to secure: Huawei’s Ascend 950DT is quoted above RMB 250,000, up 20%–50% in two months, while Cambricon’s planned 690 is reportedly up...
How Cambricon Brought DeepSeek-V4.1-Flash to vLLM on Day 0
18 sources0Cambricon says it enabled DeepSeek V4.1 Flash on the open source vLLM stack on the model’s release day by combining its NeuWare runtime, Torch MLU Ops support and optimized BangC kernels.
MiniCPM5-2B: A Small Open Model Built for Local Coding and Tool Use
16 sources0MiniCPM5 2B is a 2.52B parameter dense model aimed at local coding, tool calling and agentic workloads, with an Apache 2.0 licence.
Puffin-World: How an Open Model Combines Gravity, Depth and RGB for 3D World Modeling
16 sources0Puffin World is an open, camera centric world model that jointly represents gravity aware physics, dense depth and RGB appearance rather than predicting video pixels alone.
Unitree UnifoLM-WLA-1.0: What the 6B Humanoid Robot Model Disclosure Means
23 sources0On September 10, Unitree disclosed UnifoLM WLA 1.0, a 6B humanoid vision language action model trained on roughly 2,500 hours of real robot data to handle 64 tabletop and whole body tasks with one model.
Doubao Work Adds Parallel Agents and Mac GUI Control: Where the Speed Gains Are Real
18 sources0On September 2, 2026, Doubao Work added parallel multi agent execution and expanded its “Operate Computer” capability to macOS.
CIOE 2026: Why 800G Still Leads AI Optics as 1.6T Ramps
22 sources0800G remains the main 2026 shipment volume for AI data center optics, while 1.6T is in volume ramp up at leading suppliers but requires matching switch and server platforms, full system qualification and con...
LINXAI D30’s 33 kg Stair Climb: What the 73-Floor Test Actually Proves
14 sources0LINXAI’s D30 reportedly carried a continuous 33 kg payload up 73 floors and more than 2,000 unmodified evacuation stair steps, including six refuge floor transfers.