Building custom data-center CPUs for AI inference, targeting TikTok-scale agent workloads

ByteDance open-sources BAGEL 7B multimodal model under Apache 2.0 license

Alibaba and Nanjing University published paper claiming 9.36X speedup for million-token prefill

Released GenLIP, a generative pretraining framework for Vision Transformers.

Launched first external partnership for Qwen AI app with China Eastern Airlines for flight booking integration

Released Qwen3.6-27B, a dense 27B parameter model that surpasses its previous 397B MoE model on coding benchmarks.

Researchers developed DCW (Diffusion Correction in Wavelet domain), a method to fix SNR-t misalignment bias in diffusion models, improving performance for models like FLUX and EDM.

Moved Qwen 3.6 Plus model to API-only access, adopting a paid frontier model strategy.

Released Qwen 3.6, a free, open-weights AI model for local deployment and coding

ByteDance introduced OmniShow, a unified multimodal framework for video generation.