Deepseek drops another HUGE breakthrough

DeepSeek's speculative decoding system that builds on a parallel drafter, adding a lightweight Markov head to fix suffix decay, a confidence head to cut drafts early, and a hardware-aware algorithm that adapts draft length to GPU load. It reports 60-85% faster generation versus MTP and almost 700% higher system output under a 120 tokens/s per-user constraint, without quality loss. Code is released under the MIT license.

EVElevenLabs V4ElevenLabsAvailable in EU

The latest DeepSeek V4 model with the DSpark mechanism already implemented, downloadable from Hugging Face. The creator says it should run faster and more efficiently than the previous version, which is already fast and cheap.