DeepSeek Breaks the "Impossible Trinity" of Large Models: Stronger, Faster, and Cheaper

Insight: Beating AI Flash News – DeepSeek V4.1 Flash has nearly overhauled its entire architecture this time, aiming to simultaneously enhance performance, boost speed, and cut costs. Stronger: The model features 552 billion backbone parameters, plus an external Engram conditional memory module with 196 billion parameters. Pre-training utilized 45T multi-modal tokens, while post-training integrated a large volume of real agent tasks, tool environments, and failure cases. DeepSWE v1.1 scores 74.2%, surpassing Claude Opus 5 and GPT-5.6 Sol. Faster: The new CED architecture splits the 40-layer model into two halves. When reading prompts, only 8 billion parameters are activated per token, and just 16 billion during generation. Coupled with CSA2 cross-layer reuse and DSpark speculative decoding, the context window has been expanded from 4K to 1M (256x longer), with decoding computation per token increasing by only roughly a quarter. Cheaper: DeepSeek has further optimized KV cache to the maximum. The main cache is switched to FP4, plus cross-layer reuse, leaving global KV for each token at just 890 bytes, about 1/4 of V4 Flash; cache stored long-term on SSD or memory is further reduced to approximately 1/8. V4.1 Flash does not equate "stronger" simply to "more computation". While model scale continues to grow, only a small portion of parameters are adjusted each time; the context window is extended, yet the cache is compressed even smaller. Performance has improved, with speed and costs remaining uncompromised.