DeepSeek V3, released in December 2024, was the efficiency shock of its generation: a 671B-parameter mixture-of-experts model with 37B active per token, trained for a reported ~$5.6M in compute — one to two orders of magnitude less than Western frontier budgets — yet competitive with GPT-4o and Claude 3.5 Sonnet on standard benchmarks.

Released as open weights under a permissive license, V3 introduced multi-head latent attention and aggressive FP8 training to the open ecosystem and became the base model for DeepSeek R1 a month later. Together the pair triggered a global repricing of AI training costs in early 2025.