DSpark stage
×Stage · DeepSeek-V4.1-Flash
The speculative-decoding module is trained in a stage of its own after pre-training, with the backbone frozen.
- Trainable
- DSpark alone
- DSpark
- 3 blocks, window 128, 5 draft positions
Not disclosed Objective · Data · Steps · Learning rate · Tokens
Why
DSpark is trained after pre-training and kept training through post-training, so that it stays aligned with the policy as that changes. DeepSeek-V4.1 report §2.4.3, p. 14
Architecture at this stageThe backbone is frozen; DSpark, the speculative-decoding module, is the only part that trains.
DSpark exists from this stage on; the pinned checkpoint holds its blocks and heads.
All settings2
- Trainable
- DSpark alone; the backbone is frozen
- DSpark
- 3 Transformer blocks with a sliding window of 128 tokens; logits for 5 draft positions in one pass; a Markov head and a confidence head
Sources2
- DeepSeek_V41_Tech_Report.pdf, p. 14Trainable, DSpark
- DeepSeek_V41_Tech_Report.pdf, p. 8Trainable