Model
DeepSeek DSpark speeds up V4 inference by 60-85%
Speculative decoding framework released June 27 boosts per-user generation on V4-Flash by 60-85% and V4-Pro by 57-78% at matched throughput; open-source MIT license on HuggingFace.
Speculative decoding framework released June 27 boosts per-user generation on V4-Flash by 60-85% and V4-Pro by 57-78% at matched throughput; open-source MIT license on HuggingFace.
- Impact
- high
- Occurred
- Jun 29, 2026
- Detected
- just now
- Source
- official