DeepSeek has launched the V4.1-Flash model, a native multimodal neural network based on the Mixture of Experts architecture. The model features 552 billion parameters, with 8 billion active parameters for input and 16 billion for output. It utilizes a Causal Encoder-Decoder architecture and requires significantly less memory, using 1/4 of the HBM volume and 1/8 of the SSD storage compared to previous versions.
The new model is already available via the company's API. DeepSeek has initiated a transition process, where requests to the V4-Pro model will be automatically routed to V4.1-Flash starting September 14, 2024. The company has also announced plans to collaborate with the open-source community to support inference and explore new deployment options for the model.
Regarding corporate developments, DeepSeek has engaged CITIC Securities to prepare for an IPO on the STAR Market of the Shanghai Stock Exchange. The company aims to initiate the listing process by the end of 2026. While specific valuation and offering details remain undisclosed, the firm previously raised $7.4 billion in June, reaching a valuation exceeding $50 billion.
The company's investor base includes founder Liang Wenfeng, Tencent, and CATL. Reports suggest that a subsequent funding round could potentially value the business at approximately ¥500 billion, or roughly $75 billion. Capital raised is intended for infrastructure expansion, model development, and talent retention to compete with major industry players.
