AI
DeepSeek launches V4.1-Flash with asymmetric compute and a coming API switchover
DeepSeek's V4.1-Flash uses different compute budgets for reading prompts and generating answers, an architecture aimed at agents that consume much more context than they produce. AlphaSignal reports native vision, MIT-licensed weights and a smaller key-value cache than the previous generation. DeepSeek says the model outperforms its own V4-Pro, but independent evaluations remain limited. Existing V4-Pro API traffic is scheduled to move to Flash pricing and behavior on September 14, making compatibility testing a practical concern.
Editorial summary · Publisher article
What matters
- The reported 552-billion-parameter mixture-of-experts design activates 8 billion parameters per input token and 16 billion per output token.
- Company comparisons put cache memory at one-quarter of the previous generation; those figures are not independent workload benchmarks.
Why this made the edition
Substantive evidence supports this distinct development; selected for reader relevance and importance after comparison with published history.
Source material