DeepSeek V3
*Base-model generation preceding R1, cited in pricing shocks and replication attempts on MoE architectures.*
Shipped by DeepSeek · · class: open-weight-llm
DeepSeek V3 arrived with strong reported efficiency numbers and enough open detail to spark weeks of practitioner triage: some groups reproduced useful behavior; others emphasized hardware unknowns and evaluation caveats. The market reaction—compression in API margins—was immediate.
Subsequent months looked less like a one-shot miracle and more like a forcing function: incumbents accelerate small-model excellence, discounts, and on-prem offers when a credible low-cost anchor appears.