DeepSeek releases DSpark — V4 runs 57-85% faster via speculative decoding

DSpark is a speculative-decoding method DeepSeek says runs DeepSeek-V4 57-85% faster than its prior MTP-1 approach — pure throughput gains with no new parameters. Speed figures are vendor-reported.

See full DeepSeek profile on /llms →

Source

marktechpost.com →See it on the timeline →

ShareOG image