Specialist distillation
×Stage · DeepSeek-V3.2-Exp
Specialists trained with large-scale RL generate the domain data the final checkpoint learns from.
- Specialists
- One per domain, each fine-tuned from the same pre-trained V3.2 base and trained with large-scale RL: mathematics, competitive programming, general logical reasoning, agentic coding, agentic search, plus writing and general question answering
- Data
- The specialists’ outputs; separate generators for thinking and non-thinking responses
- Attention
- Sparse, as in the sparse stage
Not disclosed Data size, steps, settings
Sources1
- DeepSeek-V3.2 report, p. 3Specialists, Data, Attention
Key
- stage
- tokens, log scale
- a total no source states, computed from steps × batch × length
- checkpoint
- checkpoint, weights published
- a join the sources leave unexplained
- stages of a released model, folded into one row
- derived
- computed from reported values; the arithmetic is shown
- not disclosed
- no pinned source states it; nothing is filled in