DeepSeek's flash model cuts AI memory needs, hits Korean memory stocks
DeepSeek announced V4.1-Flash, a 552-billion-parameter model that activates 8 billion parameters for input and 16 billion for output, supports a one-million-token context window and reads images natively. Its architecture reuses cache across layers, stores the KV cache at 4-bit precision and reconstructs values on demand, producing about 890 bytes of cache per token and a 437-fold decline in memory per token since January 2024. The release and accompanying pricing shifts led to selling pressure in Seoul, where Samsung and SK Hynix dropped after investors priced lower per-unit memory demand for AI workloads.
V4.1-Flash reduces cache to about 890 bytes per token.
Context
DeepSeek has cut KV-cache requirements in several releases since January 2024. This change lowers the memory needed to run long AI sessions and can reduce costs. Micron reports earnings on September 30 and HBM demand projections may be…
The full analysis
19 dimensions on this story — world impact, market read, and what happens next.
- Full ContextLocked
- Affected SectorsLocked
- Stock ImpactLocked
- Economic IndicatorLocked
- Investor RelevanceLocked
- Professional RelevanceLocked
- Watch PointsLocked
- Probability of ChangeLocked
- Debate PointsLocked
- Historical ParallelLocked
- Prerequisite KnowledgeLocked
- Follow-up QuestionsLocked
- Pros & ConsLocked