DeepSeek 4.1: 256-Expert MoE, MTP-4x and DualPipe 2.0
O mercado global de inteligência artificial acaba de ser sacudido por um terremoto arquitetural de proporções históricas. A DeepSeek-AI disponibilizou oficialmente em escala global os pesos abertos e o endpoint corporativo do DeepSeek 4.1 (também denominado sob a série DeepSeek-V4.1 Foundation), consolidando uma ruptura definitiva com os paradigmas convencionais de escala densa e redefinindo os…
The global artificial intelligence market has been shaken by the release of DeepSeek 4.1, a large language model developed by DeepSeek-AI. The model has 670 billion parameters, but only uses 37 billion parameters per generated token, and has achieved parity with proprietary models from Anthropic and OpenAI. DeepSeek 4.1 can generate over 240 tokens per second per node at a cost of $0.55 per million generated tokens, significantly lower than its competitors.
The model's architecture includes a Mixture of Experts topology with 256 specialists and a shared expert, as well as a new communication scheduler called DualPipe 2.0.
Written by urgent.news from Dev.to's report — not a translation of it. Machine-written — may contain errors; check the original before relying on it.