Trending AI Posts & Hot Research Catchup - Page 48
Alibaba Cloud’s Agentic Cloud: How AgentCore, CPFS and HPN 8.0 Pro Fit Together
17 sources0At Apsara 2026, Alibaba Cloud mapped Model, Harness and Context to AI Native Cloud, Agent Native Cloud and Context Engine.
Alibaba’s Apsara 2026 Audio Models: Live Translation and the Qwen-Audio-3.1 Lineup
16 sources0Alibaba highlighted Qwen3.8 LiveTranslate and five Qwen Audio 3.1 models at Apsara 2026.
ZCode’s Repository Upload Controversy: What Zhipu Fixed and What the Audits Found
22 sources0After a researcher found ZCode uploading repository snapshots, including Git history, Zhipu open sourced the tool and removed the Repo Wiki upload workflow in v3.14.0.
Huawei OceanStor M900: A 64 PB Shared KV Cache for AI Inference
19 sources0Huawei’s OceanStor M900 extends SuperPoD KV caching from chip memory and DRAM onto shared SSD storage, with a claimed 64 PB of capacity per cluster.
Xiaomi HySparse2: How Shared KV Caches Cut Million-Token Prefill
14 sources0In a reported 80B total, roughly 3B active MoE comparison, HySparse2 uses 5.02× fewer prefill FLOPs and 2.69 GB rather than 12.09 GB of KV cache at 1 million tokens versus Hybrid SWA.
Alibaba’s Qwen Intelligence: How Its Phone Agents Work and Why Honor Is First
25 sources0Qwen Intelligence is Alibaba’s software platform for phone makers, not a new Alibaba phone.
Qwen Book: How Alibaba’s “OS as Harness” AI Computer Is Meant to Work
17 sources0Qwen Book is Alibaba Cloud’s proposed AI agent computer, unveiled in September 2026.
Huawei Atlas 960E SuperPoD: How UnifiedBus and Hi-ONE Connect 4,096 NPUs
20 sources0Announced in Shanghai on September 17, 2026, the Atlas 960E SuperPoD is Huawei’s near packaged optics AI system for up to 4,096 NPUs, rated at 8 EFLOPS FP8 or 16 EFLOPS FP4.
DeepSeek DSec: How Its Sandboxes Scale Agent Training
17 sources0DeepSeek DSec runs isolated agent actions for reinforcement learning rollouts and evaluation.
SenseNova U1 Pro: SenseTime’s 8K Image Model, Availability and Open Questions
18 sources0SenseTime launched the production SenseNova U1 Pro on September 21, 2026, through its Raccoon app and SenseNova API.
iFLYTEK Spark-ASR-2.0 vs. 1.0: Clearer Transcripts at a Claimed 10% Higher Inference Cost
18 sources0Released September 23, 2026, Spark ASR 2.0 is claimed to reduce transcription errors and produce more fluent text than Spark ASR 1.0 for roughly 10% higher inference cost.
XFEON’s Physical AI Stack: What the Product Matrix Shows—and What It Doesn’t
15 sources0On September 22, 2026, Chengdu based XFEON presented a physical AI lineup spanning chips, Token workstations, a data loop and an embodied AI model.