Trending AI Posts & Hot Research Catchup - Page 45
How a Sliding Ferroelectric Junction Turned Weak Polarization Into a 19-Million-Fold Readout
14 sources0Researchers used h BN and monolayer graphene to amplify the readout from sliding ferroelectric 3R MoS₂.
Praxis One Puts Robot AI and Motion Control on One Chip
20 sources0Unveiled in Hangzhou on September 17, 2026, Praxis One uses a MediaTek Genio Pro 5100 to bring AI decision making and motion control computing onto one chip.
DeepSeek-V4.1-Flash Explained: 1M-Token Context and KV Cache Compression
17 sources0DeepSeek V4.1 Flash supports up to one million tokens of context while activating 8B parameters per token during prefill and 16B during decode.
ZGCM-1-7B: An Open Baseline for Math Reasoning and Agentic Search
20 sources0ZGCM 1 7B is a 7.39B parameter, 256K context model released for mathematical reasoning and tool assisted search.
Huawei Cloud’s Agentic Cloud: What the Enterprise AI Stack Includes
19 sources0On September 18, 2026, Huawei Cloud presented Agentic Cloud as a stack for running enterprise AI agents—not just supplying chips.
Open-RAIL: The Software Layer That Keeps AI-Powered Robots Moving
20 sources0Open RAIL lets a robot execute one batch of model generated actions while the next is being computed, then smooths the transition between batches.
How Open-RAIL Keeps Robot Control Moving While AI Models Think
18 sources0Open RAIL sits between vision language action (VLA) or world action (WAM) models and robot hardware, so teams can change either side without rebuilding the entire execution stack.
HarmonyOS 7 Turns Xiaoyi Into an Agent That Can Work Across Apps
19 sources0HarmonyOS 7 gives Xiaoyi a system level role: it can use authorized context to plan tasks and coordinate 2,100 plus system capabilities with 2,000 plus agents and 500 plus partner Skills.
Step 5 Preview: Specs, Open-Weight Plans, Pricing and Flash Model Differences
19 sources0Step 5 Preview is a hosted, 600B parameter sparse MoE model with 27B active parameters and a 1M token context window.
Huawei’s Ascend Ecosystem Bet: Five Shifts for Long-Running AI Agents
14 sources0At HUAWEI CONNECT 2026, Zhu Zhaosheng argued that Ascend’s software community and shared 10k NPU scale compute had made it a platform others could build on.
Spirit AI’s Moz1 at CATL: What the Battery-Line Robot Does—and What Its 2027 Forecast Means
16 sources0Spirit AI says Moz1 robots are deployed at CATL and JD.com. It reports more than 99% success inserting test connectors on CATL’s battery PACK line, but that task specific figure does not establish reliabilit...
KunTai’s four-server lineup: Which system is built for which workload?
19 sources0At HUAWEI CONNECT 2026, Digital China’s KunTai introduced a Kunpeng 950 general purpose server and three Ascend 950DT AI systems.