ByteDance Plans a Muse-Style AI Assistant, Sees No Frontier Parity Before 2029
Zhang Yiming told ByteDance’s AI staff that Seed’s pretraining data is orders of magnitude smaller than that of its top US rivals, according to Bloomberg.
Zhang Yiming told ByteDance’s AI staff that Seed’s pretraining data is orders of magnitude smaller than that of its top US rivals, according to Bloomberg.
Northwestern Polytechnical University paid 874,700 yuan for two H200s in October and rented servers with H200, A100 and A800 chips, Reuters found.
The first Qwen4 model reads text, images and video across a million-token context and runs only through Alibaba’s Token Plan until November 25.
The city’s three-year robot plan adds a 50,000 PFlops training cluster and one-on-one listing help for embodied robot makers, with targets running to 2028.
The 280-billion-parameter model activates 14 billion per token, reads up to a million tokens and is already sold through Baidu’s Qianfan cloud platform.
A new Da Vinci neural processor and low-bit quantization let the phone chips run a mixture-of-experts model with about 2 billion parameters active at once.
The so.qq.com site draws on a hundred-billion-scale index spanning the open web, WeChat articles, Channels videos and academic papers.
CINNO Research counted 62 projects in the quarter, with 71.6% of their value at the planning stage and 13.6% delivered or in operation.
H3-Image, a 33.1-billion-parameter model, is pitched as producing a 4,096-by-4,096 image in under 10 seconds from up to 16 reference images.
Richard Yu named the sub-brand on November 11 while posting from an OSTA O1 Pro, and its first phones will appear at a 2:30 p.m. event shared with a car launch.