| tl;dr: DeepSeek announced DeepSeek-V4.1-Flash, the smallest model in a new architecture family featuring native visual understanding. The release emphasizes faster inference, higher throughput, and improved efficiency, and is positioned as a foundation for scaling to larger models. |
| HN Discussion:- ↑DeepSeek's transparent, detail-rich tech reports contrast favorably with Western labs' safety-focused documentation
- ↑DeepSeek is fearlessly innovative, consistently packing releases with clever ideas at frontier scale
- ~The model is no longer truly 'flash' given its increased 552B size, making local runs harder
- •Extremely low cache hit pricing may make network transfer costs dominate, reshaping API economics
- ↑Practical testing shows strong performance, willingness to comply, and interesting benchmark results
|