CXMT says it has begun mass production of JEDEC standard LPDDR6 using 16Gb dies and packages up to 16GB, with a 12,800Mbps peak rate.
Qwen Drive 1.0 4B is a 4B parameter, Apache 2.0 open weight driving VLM built on Qwen3.5 4B.
Tmall’s AI Space Station made Chinese AI capacity purchasable like prepaid credit: shoppers could buy token top ups, coding subscriptions and usage based plans through Taobao/Tmall.
Hyper3D WorldGen converts one RGB scene image into a 3D environment made of independent, editable objects rather than one fused mesh.
BitInfinite and Wuhan University reportedly finished first among 28 entrants in the ECCV 2026 CAD Challenge with a 96.39 score, on a task requiring normalized STEP BRep output from renders and technical draw...
DeepSeek V4.1 Flash was a two day API preview beginning September 8, 2026, but it became an official model on September 10 under deepseek flash.
CXMT reportedly began small quantity HBM3E risk production and supplied early samples to Alibaba’s T Head and Cambricon for testing.
Unitree’s September 7, 2026 G1 sparring clip is credible visual evidence of short horizon reactive whole body control in a contact task, but not proof of general humanoid autonomy: X2 1.0 has no published pa...
Xiaomi’s XRing O3 entering mass production turns its in house AI chip strategy into a consumer product: a 3 nm phone SoC designed to support MiMo on device AI.
AgiBot reports that scaling GE Act 2.0’s embodied training data from 300 to 30,000 hours raised zero shot real robot success from 17.1% to 44.1% on G1 OP and from 13.4% to 31.1% on G2 90D.
Ling 3.0 flash VL is an MIT licensed open weight model from Ant Group’s inclusionAI that accepts text, images, and video.
Huawei says the Kirin 9050 Pro is the first commercial use of LogicFolding, which places selected logic in vertical layers to shorten critical signal paths rather than relying only on smaller lateral features.