DeepSeek V4.1-Flash drops with native vision and a big price cut

DeepSeek 发布 V4.1-Flash：新架构带来原生视觉理解与大幅降价

DeepSeek released V4.1-Flash with a new Causal Encoder-Decoder architecture and native vision understanding — no separate vision model needed. It's a 552B MoE, activating 8B params for input and 16B for output. The post doesn't disclose the exact price cut or benchmark numbers, so I'd wait for third-party evals before getting excited.

Why it matters: DeepSeek ships a new model with a causal encoder-decoder architecture that replaces bolt-on vision components. 552B total params, only 8B/16B active during inference. The post claims a price cut but gives no specific numbers or benchmarks, so the score stays below 85 until thi...