DeepSeek V3 preceded R1 and proved Chinese labs could ship competitive open-weight MoE models at scale. Its MITMassachusetts Institute of Technology — a research university central to many computing breakthroughs.-licensed weights and transparent training details made it a default choice for self-hosted inferenceRunning a trained model to produce predictions — as opposed to the training phase that learns weights. — and a wake-up call on AI compute economics before the R1 reasoningStep-by-step logical thinking in AI models — chain-of-thought before answering hard problems. shock in January.

DeepSeek V3 showed open MoE models could rival closed APIs before the R1 reasoningStep-by-step logical thinking in AI models — chain-of-thought before answering hard problems. wave.

It shifted assumptions about who could afford frontier-scale training and release open weights.