DeepSeek V4.1 Flash rumored: 552B MoE with asymmetric CED architecture, beats flagship

_AndrewZhao · x · 2026-09-10

According to a leak by @sheriyuo, DeepSeek V4.1 Flash is a 552B-parameter MoE model with a novel Causal-Encoder-Decoder architecture featuring asymmetric activations — only 8B active for input, 16B for output — making it significantly cheaper than similar-sized models.

It reportedly uses a new pretraining recipe and heavier RL post-training, and in benchmarks surpasses flagship models including DeepSeek V4 Pro. (Unverified at time of posting; later corroborated by official release.)

Related event: Leak: DeepSeek V4.1 Flash Is 552B Asymmetric MoE(2 posts)→

Original post →

More from Models

Models channel →