Chinese AI company DeepSeek just dropped its new DeepSeek‑V4.1‑Flash model, which comes with native multimodality and kicks off a phased retirement of the V4‑Pro version. According to DeepSeek, this release is all about "expanding capabilities, speeding up output, boosting throughput, and scaling to even bigger models." V4.1‑Flash is the most compact model in their latest architecture lineup.

What is DeepSeek‑V4.1‑Flash?

V4.1‑Flash is a Mixture of Experts neural network packing a massive 552 billion parameters, built on a Causal Encoder‑Decoder architecture. During inference, only 8 billion parameters are active at the input and 16 billion at the output, striking a balance between efficiency and performance. The model’s native multimodality means it can handle multiple data types right out of the box.

Phasing Out V4‑Pro

At the same time, DeepSeek is starting a gradual phase-out of V4‑Pro. While they’re not sharing a roadmap or timeline yet, the company says the series is all about scaling up and maximizing throughput.

Heading Toward a Shanghai IPO

Alongside this tech release, DeepSeek announced it’s moving toward an IPO in Shanghai. They haven’t disclosed any further details about the potential listing.