Chinese AI company DeepSeek unveiled the DeepSeek‑V4.1‑Flash, the smallest model in its new architecture family, according to its website Thursday.
The model outperforms the earlier V4 Pro, offering higher capability and throughput while scaling to larger workloads.
The company's technical report shows a Mixture‑of‑Experts model with 552B backbone parameters, and a 1M‑token context window.
The DeepSeek‑V4.1‑Flash release coincides with DeepSeek's preparations for a STAR Market IPO.