Matthew Berman just dropped a solid breakdown of the new Deepseek V4.1 Flash model. If you run heavily loaded apps or APIs, you need to see this.

He digs into their Mixture-of-Experts (MoE) architecture and shows how they managed to massively reduce memory requirements while keeping pricing dirt cheap compared to other frontier models.
The video also tests its raw performance in practical coding and simulations. Looks like a massive win for cost-efficient deployments.
Thoughts on hosting this?