On February 13, DeepSeek started testing a brand new long-context mannequin supporting 1 million tokens in its internet and app variations, whereas its API service stays at V3.2 with 128K context.
Trade observers speculate that DeepSeek could unveil a significant new launch in the course of the upcoming Lunar New Yr, doubtlessly replicating the breakout momentum it achieved final yr.
On January 12, DeepSeek revealed a brand new analysis paper titled "Conditional Reminiscence by way of Scalable Lookup: A New Axis of Sparsity for Giant Language Fashions." Liang Wenfeng seems among the many authors.
The paper introduces "conditional reminiscence," separating static sample storage from dynamic computation by way of an Engram module. Beneath an identical parameter counts and FLOPs constraints, the strategy considerably outperforms pure MoE baseline fashions.
On December 1 final yr, DeepSeek launched two official fashions: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. V3.2 reportedly reached GPT-5-level efficiency on public reasoning benchmarks, whereas V3.2-Speciale gained gold medals at IMO 2025, CMO 2025, ICPC World Finals 2025, and IOI 2025.
NetEase Youdao Dictionary named "deepseek" its 2025 Phrase of the Yr, citing 8,672,940 annual searches. In keeping with the corporate, search curiosity surged sharply all year long, initially pushed by DeepSeek's "low-cost" breakthrough in compute effectivity and bolstered by every main product replace.
Supply：The Paper
Elevate your perspective with NextTech Information, the place innovation meets perception.
Uncover the newest breakthroughs, get unique updates, and join with a world community of future-focused thinkers.
Unlock tomorrow’s developments in the present day: learn extra, subscribe to our publication, and turn out to be a part of the NextTech group at NextTech-news.com
