MiMo V2.5 And DeepSeek V4 Prices Drop Permanently With Efficiency Gains
Reactions from ranked influencers
2 posts
some data if it's helpful: 1. mimo v2.5 and dsv4 pricing dropped (other models are official provider prices afaik) 2. with the size, the model is already quite cheap honestly. i think the issue is the nb of reasoning tokens per task, which will probably need K3.1 to improve (might be more than 3 months away tho!). good example is how K2.7 code is muchhh more efficient than K2.5 3. new arch (linear attention (KDA), latentMoE) means more room to optimize usually, but KDA has been around since kimi linear, and i'd say current model are very good at optimizing kernels so they might already have near optimal ones at release lol
@theo @ClementDelangue other models official provider prices are still the same afaik*