DeepSeek V2
A large-scale Mixture-of-Experts (MoE) language model released by DeepSeek, featuring innovations in MoE load-balancing mechanisms. It is regarded as a representative model in DeepSeek's series of efficient large language models.
Core Facts
Timeline (last 90 days)
DeepSeek V2, released in May 2024, adopted the self-developed MLA attention mechanism, cutting inference costs to one-tenth of the original.
On the prefill sizes of DeepSeek V2 and V3, FlyDSL Flex Attention achieved close to a 2x improvement.
The domestic large model API price war began with the release of DeepSeek V2 in May 2024, whose API pricing was significantly below the industry average and triggered follow-on price cuts from ByteDance, Alibaba, Baidu, and other vendors.
All Facts (4)
DeepSeek has undergone iterative model updates from V2 to V3 to R1, with both model capabilities and interaction methods changing rapidly.
65%UnverifiedDeepSeek V2, released in May 2024, adopted the self-developed MLA attention mechanism, cutting inference costs to one-tenth of the original.
50%UnverifiedOn the prefill sizes of DeepSeek V2 and V3, FlyDSL Flex Attention achieved close to a 2x improvement.
50%UnverifiedThe domestic large model API price war began with the release of DeepSeek V2 in May 2024, whose API pricing was significantly below the industry average and triggered follow-on price cuts from ByteDance, Alibaba, Baidu, and other vendors.
50%