Released V4-Pro and V4-Flash open-weight models with up to 1.6 trillion parameters

Released DeepSeek-V4 with 1M-token context at 10% KV cache cost

Initiates first external fundraising round, targeting at least $300M at a $10B+ valuation to set a benchmark for employee stock options.

Delays to DeepSeek-V4 model due to Huawei Ascend chip compatibility work

Lost core researchers Luo Fuli to Xiaomi and Guo Daya to ByteDance

Seeking first external funding round of $300M+ at $10B+ valuation

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

Introduces Helios, a 14B parameter video generation model running at 19.5 FPS on a single H100 GPU.

Collaborated with Tsinghua University and Peking University to develop the HACPO research framework.

Introduced Mixture-of-Depths Attention (MoDA) for deep LLMs